[NET]: sparse warning fixes
[linux-2.6.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks.
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/mm.h>
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained DSACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
108
109 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
114
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
116
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
119
120 /* Adapt the MSS value used to make delayed ack decision to the
121  * real world.
122  */
123 static void tcp_measure_rcv_mss(struct sock *sk,
124                                 const struct sk_buff *skb)
125 {
126         struct inet_connection_sock *icsk = inet_csk(sk);
127         const unsigned int lss = icsk->icsk_ack.last_seg_size;
128         unsigned int len;
129
130         icsk->icsk_ack.last_seg_size = 0;
131
132         /* skb->len may jitter because of SACKs, even if peer
133          * sends good full-sized frames.
134          */
135         len = skb_shinfo(skb)->gso_size ?: skb->len;
136         if (len >= icsk->icsk_ack.rcv_mss) {
137                 icsk->icsk_ack.rcv_mss = len;
138         } else {
139                 /* Otherwise, we make more careful check taking into account,
140                  * that SACKs block is variable.
141                  *
142                  * "len" is invariant segment length, including TCP header.
143                  */
144                 len += skb->data - skb_transport_header(skb);
145                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
146                     /* If PSH is not set, packet should be
147                      * full sized, provided peer TCP is not badly broken.
148                      * This observation (if it is correct 8)) allows
149                      * to handle super-low mtu links fairly.
150                      */
151                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
152                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
153                         /* Subtract also invariant (if peer is RFC compliant),
154                          * tcp header plus fixed timestamp option length.
155                          * Resulting "len" is MSS free of SACK jitter.
156                          */
157                         len -= tcp_sk(sk)->tcp_header_len;
158                         icsk->icsk_ack.last_seg_size = len;
159                         if (len == lss) {
160                                 icsk->icsk_ack.rcv_mss = len;
161                                 return;
162                         }
163                 }
164                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
165                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
166                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
167         }
168 }
169
170 static void tcp_incr_quickack(struct sock *sk)
171 {
172         struct inet_connection_sock *icsk = inet_csk(sk);
173         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
174
175         if (quickacks==0)
176                 quickacks=2;
177         if (quickacks > icsk->icsk_ack.quick)
178                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
179 }
180
181 void tcp_enter_quickack_mode(struct sock *sk)
182 {
183         struct inet_connection_sock *icsk = inet_csk(sk);
184         tcp_incr_quickack(sk);
185         icsk->icsk_ack.pingpong = 0;
186         icsk->icsk_ack.ato = TCP_ATO_MIN;
187 }
188
189 /* Send ACKs quickly, if "quick" count is not exhausted
190  * and the session is not interactive.
191  */
192
193 static inline int tcp_in_quickack_mode(const struct sock *sk)
194 {
195         const struct inet_connection_sock *icsk = inet_csk(sk);
196         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
197 }
198
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
200 {
201         if (tp->ecn_flags&TCP_ECN_OK)
202                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
203 }
204
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
206 {
207         if (tcp_hdr(skb)->cwr)
208                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
209 }
210
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
212 {
213         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214 }
215
216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
217 {
218         if (tp->ecn_flags&TCP_ECN_OK) {
219                 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
220                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
221                 /* Funny extension: if ECT is not set on a segment,
222                  * it is surely retransmit. It is not in ECN RFC,
223                  * but Linux follows this rule. */
224                 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
225                         tcp_enter_quickack_mode((struct sock *)tp);
226         }
227 }
228
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
230 {
231         if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
232                 tp->ecn_flags &= ~TCP_ECN_OK;
233 }
234
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
236 {
237         if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
238                 tp->ecn_flags &= ~TCP_ECN_OK;
239 }
240
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
242 {
243         if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
244                 return 1;
245         return 0;
246 }
247
248 /* Buffer size and advertised window tuning.
249  *
250  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
251  */
252
253 static void tcp_fixup_sndbuf(struct sock *sk)
254 {
255         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
256                      sizeof(struct sk_buff);
257
258         if (sk->sk_sndbuf < 3 * sndmem)
259                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
260 }
261
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
263  *
264  * All tcp_full_space() is split to two parts: "network" buffer, allocated
265  * forward and advertised in receiver window (tp->rcv_wnd) and
266  * "application buffer", required to isolate scheduling/application
267  * latencies from network.
268  * window_clamp is maximal advertised window. It can be less than
269  * tcp_full_space(), in this case tcp_full_space() - window_clamp
270  * is reserved for "application" buffer. The less window_clamp is
271  * the smoother our behaviour from viewpoint of network, but the lower
272  * throughput and the higher sensitivity of the connection to losses. 8)
273  *
274  * rcv_ssthresh is more strict window_clamp used at "slow start"
275  * phase to predict further behaviour of this connection.
276  * It is used for two goals:
277  * - to enforce header prediction at sender, even when application
278  *   requires some significant "application buffer". It is check #1.
279  * - to prevent pruning of receive queue because of misprediction
280  *   of receiver window. Check #2.
281  *
282  * The scheme does not work when sender sends good segments opening
283  * window and then starts to feed us spaghetti. But it should work
284  * in common situations. Otherwise, we have to rely on queue collapsing.
285  */
286
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
289 {
290         struct tcp_sock *tp = tcp_sk(sk);
291         /* Optimize this! */
292         int truesize = tcp_win_from_space(skb->truesize)/2;
293         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
294
295         while (tp->rcv_ssthresh <= window) {
296                 if (truesize <= skb->len)
297                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
298
299                 truesize >>= 1;
300                 window >>= 1;
301         }
302         return 0;
303 }
304
305 static void tcp_grow_window(struct sock *sk,
306                             struct sk_buff *skb)
307 {
308         struct tcp_sock *tp = tcp_sk(sk);
309
310         /* Check #1 */
311         if (tp->rcv_ssthresh < tp->window_clamp &&
312             (int)tp->rcv_ssthresh < tcp_space(sk) &&
313             !tcp_memory_pressure) {
314                 int incr;
315
316                 /* Check #2. Increase window, if skb with such overhead
317                  * will fit to rcvbuf in future.
318                  */
319                 if (tcp_win_from_space(skb->truesize) <= skb->len)
320                         incr = 2*tp->advmss;
321                 else
322                         incr = __tcp_grow_window(sk, skb);
323
324                 if (incr) {
325                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
326                         inet_csk(sk)->icsk_ack.quick |= 1;
327                 }
328         }
329 }
330
331 /* 3. Tuning rcvbuf, when connection enters established state. */
332
333 static void tcp_fixup_rcvbuf(struct sock *sk)
334 {
335         struct tcp_sock *tp = tcp_sk(sk);
336         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
337
338         /* Try to select rcvbuf so that 4 mss-sized segments
339          * will fit to window and corresponding skbs will fit to our rcvbuf.
340          * (was 3; 4 is minimum to allow fast retransmit to work.)
341          */
342         while (tcp_win_from_space(rcvmem) < tp->advmss)
343                 rcvmem += 128;
344         if (sk->sk_rcvbuf < 4 * rcvmem)
345                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
346 }
347
348 /* 4. Try to fixup all. It is made immediately after connection enters
349  *    established state.
350  */
351 static void tcp_init_buffer_space(struct sock *sk)
352 {
353         struct tcp_sock *tp = tcp_sk(sk);
354         int maxwin;
355
356         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
357                 tcp_fixup_rcvbuf(sk);
358         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
359                 tcp_fixup_sndbuf(sk);
360
361         tp->rcvq_space.space = tp->rcv_wnd;
362
363         maxwin = tcp_full_space(sk);
364
365         if (tp->window_clamp >= maxwin) {
366                 tp->window_clamp = maxwin;
367
368                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
369                         tp->window_clamp = max(maxwin -
370                                                (maxwin >> sysctl_tcp_app_win),
371                                                4 * tp->advmss);
372         }
373
374         /* Force reservation of one segment. */
375         if (sysctl_tcp_app_win &&
376             tp->window_clamp > 2 * tp->advmss &&
377             tp->window_clamp + tp->advmss > maxwin)
378                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
379
380         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
381         tp->snd_cwnd_stamp = tcp_time_stamp;
382 }
383
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock *sk)
386 {
387         struct tcp_sock *tp = tcp_sk(sk);
388         struct inet_connection_sock *icsk = inet_csk(sk);
389
390         icsk->icsk_ack.quick = 0;
391
392         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
393             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
394             !tcp_memory_pressure &&
395             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
396                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
397                                     sysctl_tcp_rmem[2]);
398         }
399         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
400                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
401 }
402
403
404 /* Initialize RCV_MSS value.
405  * RCV_MSS is an our guess about MSS used by the peer.
406  * We haven't any direct information about the MSS.
407  * It's better to underestimate the RCV_MSS rather than overestimate.
408  * Overestimations make us ACKing less frequently than needed.
409  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
410  */
411 void tcp_initialize_rcv_mss(struct sock *sk)
412 {
413         struct tcp_sock *tp = tcp_sk(sk);
414         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
415
416         hint = min(hint, tp->rcv_wnd/2);
417         hint = min(hint, TCP_MIN_RCVMSS);
418         hint = max(hint, TCP_MIN_MSS);
419
420         inet_csk(sk)->icsk_ack.rcv_mss = hint;
421 }
422
423 /* Receiver "autotuning" code.
424  *
425  * The algorithm for RTT estimation w/o timestamps is based on
426  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
427  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
428  *
429  * More detail on this code can be found at
430  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
431  * though this reference is out of date.  A new paper
432  * is pending.
433  */
434 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
435 {
436         u32 new_sample = tp->rcv_rtt_est.rtt;
437         long m = sample;
438
439         if (m == 0)
440                 m = 1;
441
442         if (new_sample != 0) {
443                 /* If we sample in larger samples in the non-timestamp
444                  * case, we could grossly overestimate the RTT especially
445                  * with chatty applications or bulk transfer apps which
446                  * are stalled on filesystem I/O.
447                  *
448                  * Also, since we are only going for a minimum in the
449                  * non-timestamp case, we do not smooth things out
450                  * else with timestamps disabled convergence takes too
451                  * long.
452                  */
453                 if (!win_dep) {
454                         m -= (new_sample >> 3);
455                         new_sample += m;
456                 } else if (m < new_sample)
457                         new_sample = m << 3;
458         } else {
459                 /* No previous measure. */
460                 new_sample = m << 3;
461         }
462
463         if (tp->rcv_rtt_est.rtt != new_sample)
464                 tp->rcv_rtt_est.rtt = new_sample;
465 }
466
467 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
468 {
469         if (tp->rcv_rtt_est.time == 0)
470                 goto new_measure;
471         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
472                 return;
473         tcp_rcv_rtt_update(tp,
474                            jiffies - tp->rcv_rtt_est.time,
475                            1);
476
477 new_measure:
478         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
479         tp->rcv_rtt_est.time = tcp_time_stamp;
480 }
481
482 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
483 {
484         struct tcp_sock *tp = tcp_sk(sk);
485         if (tp->rx_opt.rcv_tsecr &&
486             (TCP_SKB_CB(skb)->end_seq -
487              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
488                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
489 }
490
491 /*
492  * This function should be called every time data is copied to user space.
493  * It calculates the appropriate TCP receive buffer space.
494  */
495 void tcp_rcv_space_adjust(struct sock *sk)
496 {
497         struct tcp_sock *tp = tcp_sk(sk);
498         int time;
499         int space;
500
501         if (tp->rcvq_space.time == 0)
502                 goto new_measure;
503
504         time = tcp_time_stamp - tp->rcvq_space.time;
505         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
506             tp->rcv_rtt_est.rtt == 0)
507                 return;
508
509         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
510
511         space = max(tp->rcvq_space.space, space);
512
513         if (tp->rcvq_space.space != space) {
514                 int rcvmem;
515
516                 tp->rcvq_space.space = space;
517
518                 if (sysctl_tcp_moderate_rcvbuf &&
519                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
520                         int new_clamp = space;
521
522                         /* Receive space grows, normalize in order to
523                          * take into account packet headers and sk_buff
524                          * structure overhead.
525                          */
526                         space /= tp->advmss;
527                         if (!space)
528                                 space = 1;
529                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
530                                   16 + sizeof(struct sk_buff));
531                         while (tcp_win_from_space(rcvmem) < tp->advmss)
532                                 rcvmem += 128;
533                         space *= rcvmem;
534                         space = min(space, sysctl_tcp_rmem[2]);
535                         if (space > sk->sk_rcvbuf) {
536                                 sk->sk_rcvbuf = space;
537
538                                 /* Make the window clamp follow along.  */
539                                 tp->window_clamp = new_clamp;
540                         }
541                 }
542         }
543
544 new_measure:
545         tp->rcvq_space.seq = tp->copied_seq;
546         tp->rcvq_space.time = tcp_time_stamp;
547 }
548
549 /* There is something which you must keep in mind when you analyze the
550  * behavior of the tp->ato delayed ack timeout interval.  When a
551  * connection starts up, we want to ack as quickly as possible.  The
552  * problem is that "good" TCP's do slow start at the beginning of data
553  * transmission.  The means that until we send the first few ACK's the
554  * sender will sit on his end and only queue most of his data, because
555  * he can only send snd_cwnd unacked packets at any given time.  For
556  * each ACK we send, he increments snd_cwnd and transmits more of his
557  * queue.  -DaveM
558  */
559 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
560 {
561         struct tcp_sock *tp = tcp_sk(sk);
562         struct inet_connection_sock *icsk = inet_csk(sk);
563         u32 now;
564
565         inet_csk_schedule_ack(sk);
566
567         tcp_measure_rcv_mss(sk, skb);
568
569         tcp_rcv_rtt_measure(tp);
570
571         now = tcp_time_stamp;
572
573         if (!icsk->icsk_ack.ato) {
574                 /* The _first_ data packet received, initialize
575                  * delayed ACK engine.
576                  */
577                 tcp_incr_quickack(sk);
578                 icsk->icsk_ack.ato = TCP_ATO_MIN;
579         } else {
580                 int m = now - icsk->icsk_ack.lrcvtime;
581
582                 if (m <= TCP_ATO_MIN/2) {
583                         /* The fastest case is the first. */
584                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
585                 } else if (m < icsk->icsk_ack.ato) {
586                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
587                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
588                                 icsk->icsk_ack.ato = icsk->icsk_rto;
589                 } else if (m > icsk->icsk_rto) {
590                         /* Too long gap. Apparently sender failed to
591                          * restart window, so that we send ACKs quickly.
592                          */
593                         tcp_incr_quickack(sk);
594                         sk_stream_mem_reclaim(sk);
595                 }
596         }
597         icsk->icsk_ack.lrcvtime = now;
598
599         TCP_ECN_check_ce(tp, skb);
600
601         if (skb->len >= 128)
602                 tcp_grow_window(sk, skb);
603 }
604
605 static u32 tcp_rto_min(struct sock *sk)
606 {
607         struct dst_entry *dst = __sk_dst_get(sk);
608         u32 rto_min = TCP_RTO_MIN;
609
610         if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
611                 rto_min = dst->metrics[RTAX_RTO_MIN-1];
612         return rto_min;
613 }
614
615 /* Called to compute a smoothed rtt estimate. The data fed to this
616  * routine either comes from timestamps, or from segments that were
617  * known _not_ to have been retransmitted [see Karn/Partridge
618  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
619  * piece by Van Jacobson.
620  * NOTE: the next three routines used to be one big routine.
621  * To save cycles in the RFC 1323 implementation it was better to break
622  * it up into three procedures. -- erics
623  */
624 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
625 {
626         struct tcp_sock *tp = tcp_sk(sk);
627         long m = mrtt; /* RTT */
628
629         /*      The following amusing code comes from Jacobson's
630          *      article in SIGCOMM '88.  Note that rtt and mdev
631          *      are scaled versions of rtt and mean deviation.
632          *      This is designed to be as fast as possible
633          *      m stands for "measurement".
634          *
635          *      On a 1990 paper the rto value is changed to:
636          *      RTO = rtt + 4 * mdev
637          *
638          * Funny. This algorithm seems to be very broken.
639          * These formulae increase RTO, when it should be decreased, increase
640          * too slowly, when it should be increased quickly, decrease too quickly
641          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
642          * does not matter how to _calculate_ it. Seems, it was trap
643          * that VJ failed to avoid. 8)
644          */
645         if (m == 0)
646                 m = 1;
647         if (tp->srtt != 0) {
648                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
649                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
650                 if (m < 0) {
651                         m = -m;         /* m is now abs(error) */
652                         m -= (tp->mdev >> 2);   /* similar update on mdev */
653                         /* This is similar to one of Eifel findings.
654                          * Eifel blocks mdev updates when rtt decreases.
655                          * This solution is a bit different: we use finer gain
656                          * for mdev in this case (alpha*beta).
657                          * Like Eifel it also prevents growth of rto,
658                          * but also it limits too fast rto decreases,
659                          * happening in pure Eifel.
660                          */
661                         if (m > 0)
662                                 m >>= 3;
663                 } else {
664                         m -= (tp->mdev >> 2);   /* similar update on mdev */
665                 }
666                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
667                 if (tp->mdev > tp->mdev_max) {
668                         tp->mdev_max = tp->mdev;
669                         if (tp->mdev_max > tp->rttvar)
670                                 tp->rttvar = tp->mdev_max;
671                 }
672                 if (after(tp->snd_una, tp->rtt_seq)) {
673                         if (tp->mdev_max < tp->rttvar)
674                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
675                         tp->rtt_seq = tp->snd_nxt;
676                         tp->mdev_max = tcp_rto_min(sk);
677                 }
678         } else {
679                 /* no previous measure. */
680                 tp->srtt = m<<3;        /* take the measured time to be rtt */
681                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
682                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
683                 tp->rtt_seq = tp->snd_nxt;
684         }
685 }
686
687 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
688  * routine referred to above.
689  */
690 static inline void tcp_set_rto(struct sock *sk)
691 {
692         const struct tcp_sock *tp = tcp_sk(sk);
693         /* Old crap is replaced with new one. 8)
694          *
695          * More seriously:
696          * 1. If rtt variance happened to be less 50msec, it is hallucination.
697          *    It cannot be less due to utterly erratic ACK generation made
698          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
699          *    to do with delayed acks, because at cwnd>2 true delack timeout
700          *    is invisible. Actually, Linux-2.4 also generates erratic
701          *    ACKs in some circumstances.
702          */
703         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
704
705         /* 2. Fixups made earlier cannot be right.
706          *    If we do not estimate RTO correctly without them,
707          *    all the algo is pure shit and should be replaced
708          *    with correct one. It is exactly, which we pretend to do.
709          */
710 }
711
712 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
713  * guarantees that rto is higher.
714  */
715 static inline void tcp_bound_rto(struct sock *sk)
716 {
717         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
718                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
719 }
720
721 /* Save metrics learned by this TCP session.
722    This function is called only, when TCP finishes successfully
723    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
724  */
725 void tcp_update_metrics(struct sock *sk)
726 {
727         struct tcp_sock *tp = tcp_sk(sk);
728         struct dst_entry *dst = __sk_dst_get(sk);
729
730         if (sysctl_tcp_nometrics_save)
731                 return;
732
733         dst_confirm(dst);
734
735         if (dst && (dst->flags&DST_HOST)) {
736                 const struct inet_connection_sock *icsk = inet_csk(sk);
737                 int m;
738
739                 if (icsk->icsk_backoff || !tp->srtt) {
740                         /* This session failed to estimate rtt. Why?
741                          * Probably, no packets returned in time.
742                          * Reset our results.
743                          */
744                         if (!(dst_metric_locked(dst, RTAX_RTT)))
745                                 dst->metrics[RTAX_RTT-1] = 0;
746                         return;
747                 }
748
749                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
750
751                 /* If newly calculated rtt larger than stored one,
752                  * store new one. Otherwise, use EWMA. Remember,
753                  * rtt overestimation is always better than underestimation.
754                  */
755                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
756                         if (m <= 0)
757                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
758                         else
759                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
760                 }
761
762                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
763                         if (m < 0)
764                                 m = -m;
765
766                         /* Scale deviation to rttvar fixed point */
767                         m >>= 1;
768                         if (m < tp->mdev)
769                                 m = tp->mdev;
770
771                         if (m >= dst_metric(dst, RTAX_RTTVAR))
772                                 dst->metrics[RTAX_RTTVAR-1] = m;
773                         else
774                                 dst->metrics[RTAX_RTTVAR-1] -=
775                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
776                 }
777
778                 if (tp->snd_ssthresh >= 0xFFFF) {
779                         /* Slow start still did not finish. */
780                         if (dst_metric(dst, RTAX_SSTHRESH) &&
781                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
782                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
783                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
784                         if (!dst_metric_locked(dst, RTAX_CWND) &&
785                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
786                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
787                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
788                            icsk->icsk_ca_state == TCP_CA_Open) {
789                         /* Cong. avoidance phase, cwnd is reliable. */
790                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
791                                 dst->metrics[RTAX_SSTHRESH-1] =
792                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
793                         if (!dst_metric_locked(dst, RTAX_CWND))
794                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
795                 } else {
796                         /* Else slow start did not finish, cwnd is non-sense,
797                            ssthresh may be also invalid.
798                          */
799                         if (!dst_metric_locked(dst, RTAX_CWND))
800                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
801                         if (dst->metrics[RTAX_SSTHRESH-1] &&
802                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
803                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
804                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
805                 }
806
807                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
808                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
809                             tp->reordering != sysctl_tcp_reordering)
810                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
811                 }
812         }
813 }
814
815 /* Numbers are taken from RFC3390.
816  *
817  * John Heffner states:
818  *
819  *      The RFC specifies a window of no more than 4380 bytes
820  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
821  *      is a bit misleading because they use a clamp at 4380 bytes
822  *      rather than use a multiplier in the relevant range.
823  */
824 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
825 {
826         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
827
828         if (!cwnd) {
829                 if (tp->mss_cache > 1460)
830                         cwnd = 2;
831                 else
832                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
833         }
834         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
835 }
836
837 /* Set slow start threshold and cwnd not falling to slow start */
838 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
839 {
840         struct tcp_sock *tp = tcp_sk(sk);
841         const struct inet_connection_sock *icsk = inet_csk(sk);
842
843         tp->prior_ssthresh = 0;
844         tp->bytes_acked = 0;
845         if (icsk->icsk_ca_state < TCP_CA_CWR) {
846                 tp->undo_marker = 0;
847                 if (set_ssthresh)
848                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
849                 tp->snd_cwnd = min(tp->snd_cwnd,
850                                    tcp_packets_in_flight(tp) + 1U);
851                 tp->snd_cwnd_cnt = 0;
852                 tp->high_seq = tp->snd_nxt;
853                 tp->snd_cwnd_stamp = tcp_time_stamp;
854                 TCP_ECN_queue_cwr(tp);
855
856                 tcp_set_ca_state(sk, TCP_CA_CWR);
857         }
858 }
859
860 /*
861  * Packet counting of FACK is based on in-order assumptions, therefore TCP
862  * disables it when reordering is detected
863  */
864 static void tcp_disable_fack(struct tcp_sock *tp)
865 {
866         tp->rx_opt.sack_ok &= ~2;
867 }
868
869 /* Take a notice that peer is sending DSACKs */
870 static void tcp_dsack_seen(struct tcp_sock *tp)
871 {
872         tp->rx_opt.sack_ok |= 4;
873 }
874
875 /* Initialize metrics on socket. */
876
877 static void tcp_init_metrics(struct sock *sk)
878 {
879         struct tcp_sock *tp = tcp_sk(sk);
880         struct dst_entry *dst = __sk_dst_get(sk);
881
882         if (dst == NULL)
883                 goto reset;
884
885         dst_confirm(dst);
886
887         if (dst_metric_locked(dst, RTAX_CWND))
888                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
889         if (dst_metric(dst, RTAX_SSTHRESH)) {
890                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
891                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
892                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
893         }
894         if (dst_metric(dst, RTAX_REORDERING) &&
895             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
896                 tcp_disable_fack(tp);
897                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
898         }
899
900         if (dst_metric(dst, RTAX_RTT) == 0)
901                 goto reset;
902
903         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
904                 goto reset;
905
906         /* Initial rtt is determined from SYN,SYN-ACK.
907          * The segment is small and rtt may appear much
908          * less than real one. Use per-dst memory
909          * to make it more realistic.
910          *
911          * A bit of theory. RTT is time passed after "normal" sized packet
912          * is sent until it is ACKed. In normal circumstances sending small
913          * packets force peer to delay ACKs and calculation is correct too.
914          * The algorithm is adaptive and, provided we follow specs, it
915          * NEVER underestimate RTT. BUT! If peer tries to make some clever
916          * tricks sort of "quick acks" for time long enough to decrease RTT
917          * to low value, and then abruptly stops to do it and starts to delay
918          * ACKs, wait for troubles.
919          */
920         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
921                 tp->srtt = dst_metric(dst, RTAX_RTT);
922                 tp->rtt_seq = tp->snd_nxt;
923         }
924         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
925                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
926                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
927         }
928         tcp_set_rto(sk);
929         tcp_bound_rto(sk);
930         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
931                 goto reset;
932         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
933         tp->snd_cwnd_stamp = tcp_time_stamp;
934         return;
935
936 reset:
937         /* Play conservative. If timestamps are not
938          * supported, TCP will fail to recalculate correct
939          * rtt, if initial rto is too small. FORGET ALL AND RESET!
940          */
941         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
942                 tp->srtt = 0;
943                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
944                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
945         }
946 }
947
948 static void tcp_update_reordering(struct sock *sk, const int metric,
949                                   const int ts)
950 {
951         struct tcp_sock *tp = tcp_sk(sk);
952         if (metric > tp->reordering) {
953                 tp->reordering = min(TCP_MAX_REORDERING, metric);
954
955                 /* This exciting event is worth to be remembered. 8) */
956                 if (ts)
957                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
958                 else if (tcp_is_reno(tp))
959                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
960                 else if (tcp_is_fack(tp))
961                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
962                 else
963                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
964 #if FASTRETRANS_DEBUG > 1
965                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
966                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
967                        tp->reordering,
968                        tp->fackets_out,
969                        tp->sacked_out,
970                        tp->undo_marker ? tp->undo_retrans : 0);
971 #endif
972                 tcp_disable_fack(tp);
973         }
974 }
975
976 /* This procedure tags the retransmission queue when SACKs arrive.
977  *
978  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
979  * Packets in queue with these bits set are counted in variables
980  * sacked_out, retrans_out and lost_out, correspondingly.
981  *
982  * Valid combinations are:
983  * Tag  InFlight        Description
984  * 0    1               - orig segment is in flight.
985  * S    0               - nothing flies, orig reached receiver.
986  * L    0               - nothing flies, orig lost by net.
987  * R    2               - both orig and retransmit are in flight.
988  * L|R  1               - orig is lost, retransmit is in flight.
989  * S|R  1               - orig reached receiver, retrans is still in flight.
990  * (L|S|R is logically valid, it could occur when L|R is sacked,
991  *  but it is equivalent to plain S and code short-curcuits it to S.
992  *  L|S is logically invalid, it would mean -1 packet in flight 8))
993  *
994  * These 6 states form finite state machine, controlled by the following events:
995  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
996  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
997  * 3. Loss detection event of one of three flavors:
998  *      A. Scoreboard estimator decided the packet is lost.
999  *         A'. Reno "three dupacks" marks head of queue lost.
1000  *         A''. Its FACK modfication, head until snd.fack is lost.
1001  *      B. SACK arrives sacking data transmitted after never retransmitted
1002  *         hole was sent out.
1003  *      C. SACK arrives sacking SND.NXT at the moment, when the
1004  *         segment was retransmitted.
1005  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1006  *
1007  * It is pleasant to note, that state diagram turns out to be commutative,
1008  * so that we are allowed not to be bothered by order of our actions,
1009  * when multiple events arrive simultaneously. (see the function below).
1010  *
1011  * Reordering detection.
1012  * --------------------
1013  * Reordering metric is maximal distance, which a packet can be displaced
1014  * in packet stream. With SACKs we can estimate it:
1015  *
1016  * 1. SACK fills old hole and the corresponding segment was not
1017  *    ever retransmitted -> reordering. Alas, we cannot use it
1018  *    when segment was retransmitted.
1019  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1020  *    for retransmitted and already SACKed segment -> reordering..
1021  * Both of these heuristics are not used in Loss state, when we cannot
1022  * account for retransmits accurately.
1023  *
1024  * SACK block validation.
1025  * ----------------------
1026  *
1027  * SACK block range validation checks that the received SACK block fits to
1028  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1029  * Note that SND.UNA is not included to the range though being valid because
1030  * it means that the receiver is rather inconsistent with itself reporting
1031  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1032  * perfectly valid, however, in light of RFC2018 which explicitly states
1033  * that "SACK block MUST reflect the newest segment.  Even if the newest
1034  * segment is going to be discarded ...", not that it looks very clever
1035  * in case of head skb. Due to potentional receiver driven attacks, we
1036  * choose to avoid immediate execution of a walk in write queue due to
1037  * reneging and defer head skb's loss recovery to standard loss recovery
1038  * procedure that will eventually trigger (nothing forbids us doing this).
1039  *
1040  * Implements also blockage to start_seq wrap-around. Problem lies in the
1041  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1042  * there's no guarantee that it will be before snd_nxt (n). The problem
1043  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1044  * wrap (s_w):
1045  *
1046  *         <- outs wnd ->                          <- wrapzone ->
1047  *         u     e      n                         u_w   e_w  s n_w
1048  *         |     |      |                          |     |   |  |
1049  * |<------------+------+----- TCP seqno space --------------+---------->|
1050  * ...-- <2^31 ->|                                           |<--------...
1051  * ...---- >2^31 ------>|                                    |<--------...
1052  *
1053  * Current code wouldn't be vulnerable but it's better still to discard such
1054  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1055  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1056  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1057  * equal to the ideal case (infinite seqno space without wrap caused issues).
1058  *
1059  * With D-SACK the lower bound is extended to cover sequence space below
1060  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1061  * again, DSACK block must not to go across snd_una (for the same reason as
1062  * for the normal SACK blocks, explained above). But there all simplicity
1063  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1064  * fully below undo_marker they do not affect behavior in anyway and can
1065  * therefore be safely ignored. In rare cases (which are more or less
1066  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1067  * fragmentation and packet reordering past skb's retransmission. To consider
1068  * them correctly, the acceptable range must be extended even more though
1069  * the exact amount is rather hard to quantify. However, tp->max_window can
1070  * be used as an exaggerated estimate.
1071  */
1072 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1073                                   u32 start_seq, u32 end_seq)
1074 {
1075         /* Too far in future, or reversed (interpretation is ambiguous) */
1076         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1077                 return 0;
1078
1079         /* Nasty start_seq wrap-around check (see comments above) */
1080         if (!before(start_seq, tp->snd_nxt))
1081                 return 0;
1082
1083         /* In outstanding window? ...This is valid exit for DSACKs too.
1084          * start_seq == snd_una is non-sensical (see comments above)
1085          */
1086         if (after(start_seq, tp->snd_una))
1087                 return 1;
1088
1089         if (!is_dsack || !tp->undo_marker)
1090                 return 0;
1091
1092         /* ...Then it's D-SACK, and must reside below snd_una completely */
1093         if (!after(end_seq, tp->snd_una))
1094                 return 0;
1095
1096         if (!before(start_seq, tp->undo_marker))
1097                 return 1;
1098
1099         /* Too old */
1100         if (!after(end_seq, tp->undo_marker))
1101                 return 0;
1102
1103         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1104          *   start_seq < undo_marker and end_seq >= undo_marker.
1105          */
1106         return !before(start_seq, end_seq - tp->max_window);
1107 }
1108
1109
1110 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1111                            struct tcp_sack_block_wire *sp, int num_sacks,
1112                            u32 prior_snd_una)
1113 {
1114         u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1115         u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1116         int dup_sack = 0;
1117
1118         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1119                 dup_sack = 1;
1120                 tcp_dsack_seen(tp);
1121                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1122         } else if (num_sacks > 1) {
1123                 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1124                 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1125
1126                 if (!after(end_seq_0, end_seq_1) &&
1127                     !before(start_seq_0, start_seq_1)) {
1128                         dup_sack = 1;
1129                         tcp_dsack_seen(tp);
1130                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1131                 }
1132         }
1133
1134         /* D-SACK for already forgotten data... Do dumb counting. */
1135         if (dup_sack &&
1136             !after(end_seq_0, prior_snd_una) &&
1137             after(end_seq_0, tp->undo_marker))
1138                 tp->undo_retrans--;
1139
1140         return dup_sack;
1141 }
1142
1143 static int
1144 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1145 {
1146         const struct inet_connection_sock *icsk = inet_csk(sk);
1147         struct tcp_sock *tp = tcp_sk(sk);
1148         unsigned char *ptr = (skb_transport_header(ack_skb) +
1149                               TCP_SKB_CB(ack_skb)->sacked);
1150         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1151         struct sk_buff *cached_skb;
1152         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1153         int reord = tp->packets_out;
1154         int prior_fackets;
1155         u32 lost_retrans = 0;
1156         int flag = 0;
1157         int found_dup_sack = 0;
1158         int cached_fack_count;
1159         int i;
1160         int first_sack_index;
1161
1162         if (!tp->sacked_out) {
1163                 if (WARN_ON(tp->fackets_out))
1164                         tp->fackets_out = 0;
1165                 tp->highest_sack = tp->snd_una;
1166         }
1167         prior_fackets = tp->fackets_out;
1168
1169         found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1170                                          num_sacks, prior_snd_una);
1171         if (found_dup_sack)
1172                 flag |= FLAG_DSACKING_ACK;
1173
1174         /* Eliminate too old ACKs, but take into
1175          * account more or less fresh ones, they can
1176          * contain valid SACK info.
1177          */
1178         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1179                 return 0;
1180
1181         /* SACK fastpath:
1182          * if the only SACK change is the increase of the end_seq of
1183          * the first block then only apply that SACK block
1184          * and use retrans queue hinting otherwise slowpath */
1185         flag = 1;
1186         for (i = 0; i < num_sacks; i++) {
1187                 __be32 start_seq = sp[i].start_seq;
1188                 __be32 end_seq = sp[i].end_seq;
1189
1190                 if (i == 0) {
1191                         if (tp->recv_sack_cache[i].start_seq != start_seq)
1192                                 flag = 0;
1193                 } else {
1194                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1195                             (tp->recv_sack_cache[i].end_seq != end_seq))
1196                                 flag = 0;
1197                 }
1198                 tp->recv_sack_cache[i].start_seq = start_seq;
1199                 tp->recv_sack_cache[i].end_seq = end_seq;
1200         }
1201         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1202         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1203                 tp->recv_sack_cache[i].start_seq = 0;
1204                 tp->recv_sack_cache[i].end_seq = 0;
1205         }
1206
1207         first_sack_index = 0;
1208         if (flag)
1209                 num_sacks = 1;
1210         else {
1211                 int j;
1212                 tp->fastpath_skb_hint = NULL;
1213
1214                 /* order SACK blocks to allow in order walk of the retrans queue */
1215                 for (i = num_sacks-1; i > 0; i--) {
1216                         for (j = 0; j < i; j++){
1217                                 if (after(ntohl(sp[j].start_seq),
1218                                           ntohl(sp[j+1].start_seq))){
1219                                         struct tcp_sack_block_wire tmp;
1220
1221                                         tmp = sp[j];
1222                                         sp[j] = sp[j+1];
1223                                         sp[j+1] = tmp;
1224
1225                                         /* Track where the first SACK block goes to */
1226                                         if (j == first_sack_index)
1227                                                 first_sack_index = j+1;
1228                                 }
1229
1230                         }
1231                 }
1232         }
1233
1234         /* clear flag as used for different purpose in following code */
1235         flag = 0;
1236
1237         /* Use SACK fastpath hint if valid */
1238         cached_skb = tp->fastpath_skb_hint;
1239         cached_fack_count = tp->fastpath_cnt_hint;
1240         if (!cached_skb) {
1241                 cached_skb = tcp_write_queue_head(sk);
1242                 cached_fack_count = 0;
1243         }
1244
1245         for (i=0; i<num_sacks; i++, sp++) {
1246                 struct sk_buff *skb;
1247                 __u32 start_seq = ntohl(sp->start_seq);
1248                 __u32 end_seq = ntohl(sp->end_seq);
1249                 int fack_count;
1250                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1251
1252                 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) {
1253                         if (dup_sack) {
1254                                 if (!tp->undo_marker)
1255                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1256                                 else
1257                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1258                         } else {
1259                                 /* Don't count olds caused by ACK reordering */
1260                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1261                                     !after(end_seq, tp->snd_una))
1262                                         continue;
1263                                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1264                         }
1265                         continue;
1266                 }
1267
1268                 skb = cached_skb;
1269                 fack_count = cached_fack_count;
1270
1271                 /* Event "B" in the comment above. */
1272                 if (after(end_seq, tp->high_seq))
1273                         flag |= FLAG_DATA_LOST;
1274
1275                 tcp_for_write_queue_from(skb, sk) {
1276                         int in_sack, pcount;
1277                         u8 sacked;
1278
1279                         if (skb == tcp_send_head(sk))
1280                                 break;
1281
1282                         cached_skb = skb;
1283                         cached_fack_count = fack_count;
1284                         if (i == first_sack_index) {
1285                                 tp->fastpath_skb_hint = skb;
1286                                 tp->fastpath_cnt_hint = fack_count;
1287                         }
1288
1289                         /* The retransmission queue is always in order, so
1290                          * we can short-circuit the walk early.
1291                          */
1292                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1293                                 break;
1294
1295                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1296                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1297
1298                         pcount = tcp_skb_pcount(skb);
1299
1300                         if (pcount > 1 && !in_sack &&
1301                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1302                                 unsigned int pkt_len;
1303
1304                                 in_sack = !after(start_seq,
1305                                                  TCP_SKB_CB(skb)->seq);
1306
1307                                 if (!in_sack)
1308                                         pkt_len = (start_seq -
1309                                                    TCP_SKB_CB(skb)->seq);
1310                                 else
1311                                         pkt_len = (end_seq -
1312                                                    TCP_SKB_CB(skb)->seq);
1313                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1314                                         break;
1315                                 pcount = tcp_skb_pcount(skb);
1316                         }
1317
1318                         fack_count += pcount;
1319
1320                         sacked = TCP_SKB_CB(skb)->sacked;
1321
1322                         /* Account D-SACK for retransmitted packet. */
1323                         if ((dup_sack && in_sack) &&
1324                             (sacked & TCPCB_RETRANS) &&
1325                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1326                                 tp->undo_retrans--;
1327
1328                         /* The frame is ACKed. */
1329                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1330                                 if (sacked&TCPCB_RETRANS) {
1331                                         if ((dup_sack && in_sack) &&
1332                                             (sacked&TCPCB_SACKED_ACKED))
1333                                                 reord = min(fack_count, reord);
1334                                 } else {
1335                                         /* If it was in a hole, we detected reordering. */
1336                                         if (fack_count < prior_fackets &&
1337                                             !(sacked&TCPCB_SACKED_ACKED))
1338                                                 reord = min(fack_count, reord);
1339                                 }
1340
1341                                 /* Nothing to do; acked frame is about to be dropped. */
1342                                 continue;
1343                         }
1344
1345                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1346                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1347                             (!lost_retrans || after(end_seq, lost_retrans)))
1348                                 lost_retrans = end_seq;
1349
1350                         if (!in_sack)
1351                                 continue;
1352
1353                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1354                                 if (sacked & TCPCB_SACKED_RETRANS) {
1355                                         /* If the segment is not tagged as lost,
1356                                          * we do not clear RETRANS, believing
1357                                          * that retransmission is still in flight.
1358                                          */
1359                                         if (sacked & TCPCB_LOST) {
1360                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1361                                                 tp->lost_out -= tcp_skb_pcount(skb);
1362                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1363
1364                                                 /* clear lost hint */
1365                                                 tp->retransmit_skb_hint = NULL;
1366                                         }
1367                                 } else {
1368                                         /* New sack for not retransmitted frame,
1369                                          * which was in hole. It is reordering.
1370                                          */
1371                                         if (!(sacked & TCPCB_RETRANS) &&
1372                                             fack_count < prior_fackets)
1373                                                 reord = min(fack_count, reord);
1374
1375                                         if (sacked & TCPCB_LOST) {
1376                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1377                                                 tp->lost_out -= tcp_skb_pcount(skb);
1378
1379                                                 /* clear lost hint */
1380                                                 tp->retransmit_skb_hint = NULL;
1381                                         }
1382                                         /* SACK enhanced F-RTO detection.
1383                                          * Set flag if and only if non-rexmitted
1384                                          * segments below frto_highmark are
1385                                          * SACKed (RFC4138; Appendix B).
1386                                          * Clearing correct due to in-order walk
1387                                          */
1388                                         if (after(end_seq, tp->frto_highmark)) {
1389                                                 flag &= ~FLAG_ONLY_ORIG_SACKED;
1390                                         } else {
1391                                                 if (!(sacked & TCPCB_RETRANS))
1392                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1393                                         }
1394                                 }
1395
1396                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1397                                 flag |= FLAG_DATA_SACKED;
1398                                 tp->sacked_out += tcp_skb_pcount(skb);
1399
1400                                 if (fack_count > tp->fackets_out)
1401                                         tp->fackets_out = fack_count;
1402
1403                                 if (after(TCP_SKB_CB(skb)->seq,
1404                                     tp->highest_sack))
1405                                         tp->highest_sack = TCP_SKB_CB(skb)->seq;
1406                         } else {
1407                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1408                                         reord = min(fack_count, reord);
1409                         }
1410
1411                         /* D-SACK. We can detect redundant retransmission
1412                          * in S|R and plain R frames and clear it.
1413                          * undo_retrans is decreased above, L|R frames
1414                          * are accounted above as well.
1415                          */
1416                         if (dup_sack &&
1417                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1418                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1419                                 tp->retrans_out -= tcp_skb_pcount(skb);
1420                                 tp->retransmit_skb_hint = NULL;
1421                         }
1422                 }
1423         }
1424
1425         /* Check for lost retransmit. This superb idea is
1426          * borrowed from "ratehalving". Event "C".
1427          * Later note: FACK people cheated me again 8),
1428          * we have to account for reordering! Ugly,
1429          * but should help.
1430          */
1431         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1432                 struct sk_buff *skb;
1433
1434                 tcp_for_write_queue(skb, sk) {
1435                         if (skb == tcp_send_head(sk))
1436                                 break;
1437                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1438                                 break;
1439                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1440                                 continue;
1441                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1442                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1443                             (tcp_is_fack(tp) ||
1444                              !before(lost_retrans,
1445                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1446                                      tp->mss_cache))) {
1447                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1448                                 tp->retrans_out -= tcp_skb_pcount(skb);
1449
1450                                 /* clear lost hint */
1451                                 tp->retransmit_skb_hint = NULL;
1452
1453                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1454                                         tp->lost_out += tcp_skb_pcount(skb);
1455                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1456                                         flag |= FLAG_DATA_SACKED;
1457                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1458                                 }
1459                         }
1460                 }
1461         }
1462
1463         tcp_verify_left_out(tp);
1464
1465         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1466             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1467                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1468
1469 #if FASTRETRANS_DEBUG > 0
1470         BUG_TRAP((int)tp->sacked_out >= 0);
1471         BUG_TRAP((int)tp->lost_out >= 0);
1472         BUG_TRAP((int)tp->retrans_out >= 0);
1473         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1474 #endif
1475         return flag;
1476 }
1477
1478 /* If we receive more dupacks than we expected counting segments
1479  * in assumption of absent reordering, interpret this as reordering.
1480  * The only another reason could be bug in receiver TCP.
1481  */
1482 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1483 {
1484         struct tcp_sock *tp = tcp_sk(sk);
1485         u32 holes;
1486
1487         holes = max(tp->lost_out, 1U);
1488         holes = min(holes, tp->packets_out);
1489
1490         if ((tp->sacked_out + holes) > tp->packets_out) {
1491                 tp->sacked_out = tp->packets_out - holes;
1492                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1493         }
1494 }
1495
1496 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1497
1498 static void tcp_add_reno_sack(struct sock *sk)
1499 {
1500         struct tcp_sock *tp = tcp_sk(sk);
1501         tp->sacked_out++;
1502         tcp_check_reno_reordering(sk, 0);
1503         tcp_verify_left_out(tp);
1504 }
1505
1506 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1507
1508 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1509 {
1510         struct tcp_sock *tp = tcp_sk(sk);
1511
1512         if (acked > 0) {
1513                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1514                 if (acked-1 >= tp->sacked_out)
1515                         tp->sacked_out = 0;
1516                 else
1517                         tp->sacked_out -= acked-1;
1518         }
1519         tcp_check_reno_reordering(sk, acked);
1520         tcp_verify_left_out(tp);
1521 }
1522
1523 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1524 {
1525         tp->sacked_out = 0;
1526 }
1527
1528 /* F-RTO can only be used if TCP has never retransmitted anything other than
1529  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1530  */
1531 int tcp_use_frto(struct sock *sk)
1532 {
1533         const struct tcp_sock *tp = tcp_sk(sk);
1534         struct sk_buff *skb;
1535
1536         if (!sysctl_tcp_frto)
1537                 return 0;
1538
1539         if (IsSackFrto())
1540                 return 1;
1541
1542         /* Avoid expensive walking of rexmit queue if possible */
1543         if (tp->retrans_out > 1)
1544                 return 0;
1545
1546         skb = tcp_write_queue_head(sk);
1547         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1548         tcp_for_write_queue_from(skb, sk) {
1549                 if (skb == tcp_send_head(sk))
1550                         break;
1551                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1552                         return 0;
1553                 /* Short-circuit when first non-SACKed skb has been checked */
1554                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1555                         break;
1556         }
1557         return 1;
1558 }
1559
1560 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1561  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1562  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1563  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1564  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1565  * bits are handled if the Loss state is really to be entered (in
1566  * tcp_enter_frto_loss).
1567  *
1568  * Do like tcp_enter_loss() would; when RTO expires the second time it
1569  * does:
1570  *  "Reduce ssthresh if it has not yet been made inside this window."
1571  */
1572 void tcp_enter_frto(struct sock *sk)
1573 {
1574         const struct inet_connection_sock *icsk = inet_csk(sk);
1575         struct tcp_sock *tp = tcp_sk(sk);
1576         struct sk_buff *skb;
1577
1578         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1579             tp->snd_una == tp->high_seq ||
1580             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1581              !icsk->icsk_retransmits)) {
1582                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1583                 /* Our state is too optimistic in ssthresh() call because cwnd
1584                  * is not reduced until tcp_enter_frto_loss() when previous FRTO
1585                  * recovery has not yet completed. Pattern would be this: RTO,
1586                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1587                  * up here twice).
1588                  * RFC4138 should be more specific on what to do, even though
1589                  * RTO is quite unlikely to occur after the first Cumulative ACK
1590                  * due to back-off and complexity of triggering events ...
1591                  */
1592                 if (tp->frto_counter) {
1593                         u32 stored_cwnd;
1594                         stored_cwnd = tp->snd_cwnd;
1595                         tp->snd_cwnd = 2;
1596                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1597                         tp->snd_cwnd = stored_cwnd;
1598                 } else {
1599                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1600                 }
1601                 /* ... in theory, cong.control module could do "any tricks" in
1602                  * ssthresh(), which means that ca_state, lost bits and lost_out
1603                  * counter would have to be faked before the call occurs. We
1604                  * consider that too expensive, unlikely and hacky, so modules
1605                  * using these in ssthresh() must deal these incompatibility
1606                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1607                  */
1608                 tcp_ca_event(sk, CA_EVENT_FRTO);
1609         }
1610
1611         tp->undo_marker = tp->snd_una;
1612         tp->undo_retrans = 0;
1613
1614         skb = tcp_write_queue_head(sk);
1615         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1616                 tp->undo_marker = 0;
1617         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1618                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1619                 tp->retrans_out -= tcp_skb_pcount(skb);
1620         }
1621         tcp_verify_left_out(tp);
1622
1623         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1624          * The last condition is necessary at least in tp->frto_counter case.
1625          */
1626         if (IsSackFrto() && (tp->frto_counter ||
1627             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1628             after(tp->high_seq, tp->snd_una)) {
1629                 tp->frto_highmark = tp->high_seq;
1630         } else {
1631                 tp->frto_highmark = tp->snd_nxt;
1632         }
1633         tcp_set_ca_state(sk, TCP_CA_Disorder);
1634         tp->high_seq = tp->snd_nxt;
1635         tp->frto_counter = 1;
1636 }
1637
1638 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1639  * which indicates that we should follow the traditional RTO recovery,
1640  * i.e. mark everything lost and do go-back-N retransmission.
1641  */
1642 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1643 {
1644         struct tcp_sock *tp = tcp_sk(sk);
1645         struct sk_buff *skb;
1646
1647         tp->lost_out = 0;
1648         tp->retrans_out = 0;
1649         if (tcp_is_reno(tp))
1650                 tcp_reset_reno_sack(tp);
1651
1652         tcp_for_write_queue(skb, sk) {
1653                 if (skb == tcp_send_head(sk))
1654                         break;
1655                 /*
1656                  * Count the retransmission made on RTO correctly (only when
1657                  * waiting for the first ACK and did not get it)...
1658                  */
1659                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1660                         /* For some reason this R-bit might get cleared? */
1661                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1662                                 tp->retrans_out += tcp_skb_pcount(skb);
1663                         /* ...enter this if branch just for the first segment */
1664                         flag |= FLAG_DATA_ACKED;
1665                 } else {
1666                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1667                                 tp->undo_marker = 0;
1668                         TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1669                 }
1670
1671                 /* Don't lost mark skbs that were fwd transmitted after RTO */
1672                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1673                     !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1674                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1675                         tp->lost_out += tcp_skb_pcount(skb);
1676                 }
1677         }
1678         tcp_verify_left_out(tp);
1679
1680         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1681         tp->snd_cwnd_cnt = 0;
1682         tp->snd_cwnd_stamp = tcp_time_stamp;
1683         tp->frto_counter = 0;
1684
1685         tp->reordering = min_t(unsigned int, tp->reordering,
1686                                              sysctl_tcp_reordering);
1687         tcp_set_ca_state(sk, TCP_CA_Loss);
1688         tp->high_seq = tp->frto_highmark;
1689         TCP_ECN_queue_cwr(tp);
1690
1691         tcp_clear_retrans_hints_partial(tp);
1692 }
1693
1694 void tcp_clear_retrans(struct tcp_sock *tp)
1695 {
1696         tp->retrans_out = 0;
1697
1698         tp->fackets_out = 0;
1699         tp->sacked_out = 0;
1700         tp->lost_out = 0;
1701
1702         tp->undo_marker = 0;
1703         tp->undo_retrans = 0;
1704 }
1705
1706 /* Enter Loss state. If "how" is not zero, forget all SACK information
1707  * and reset tags completely, otherwise preserve SACKs. If receiver
1708  * dropped its ofo queue, we will know this due to reneging detection.
1709  */
1710 void tcp_enter_loss(struct sock *sk, int how)
1711 {
1712         const struct inet_connection_sock *icsk = inet_csk(sk);
1713         struct tcp_sock *tp = tcp_sk(sk);
1714         struct sk_buff *skb;
1715         int cnt = 0;
1716
1717         /* Reduce ssthresh if it has not yet been made inside this window. */
1718         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1719             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1720                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1721                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1722                 tcp_ca_event(sk, CA_EVENT_LOSS);
1723         }
1724         tp->snd_cwnd       = 1;
1725         tp->snd_cwnd_cnt   = 0;
1726         tp->snd_cwnd_stamp = tcp_time_stamp;
1727
1728         tp->bytes_acked = 0;
1729         tcp_clear_retrans(tp);
1730
1731         if (!how) {
1732                 /* Push undo marker, if it was plain RTO and nothing
1733                  * was retransmitted. */
1734                 tp->undo_marker = tp->snd_una;
1735                 tcp_clear_retrans_hints_partial(tp);
1736         } else {
1737                 tcp_clear_all_retrans_hints(tp);
1738         }
1739
1740         tcp_for_write_queue(skb, sk) {
1741                 if (skb == tcp_send_head(sk))
1742                         break;
1743                 cnt += tcp_skb_pcount(skb);
1744                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1745                         tp->undo_marker = 0;
1746                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1747                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1748                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1749                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1750                         tp->lost_out += tcp_skb_pcount(skb);
1751                 } else {
1752                         tp->sacked_out += tcp_skb_pcount(skb);
1753                         tp->fackets_out = cnt;
1754                 }
1755         }
1756         tcp_verify_left_out(tp);
1757
1758         tp->reordering = min_t(unsigned int, tp->reordering,
1759                                              sysctl_tcp_reordering);
1760         tcp_set_ca_state(sk, TCP_CA_Loss);
1761         tp->high_seq = tp->snd_nxt;
1762         TCP_ECN_queue_cwr(tp);
1763         /* Abort FRTO algorithm if one is in progress */
1764         tp->frto_counter = 0;
1765 }
1766
1767 static int tcp_check_sack_reneging(struct sock *sk)
1768 {
1769         struct sk_buff *skb;
1770
1771         /* If ACK arrived pointing to a remembered SACK,
1772          * it means that our remembered SACKs do not reflect
1773          * real state of receiver i.e.
1774          * receiver _host_ is heavily congested (or buggy).
1775          * Do processing similar to RTO timeout.
1776          */
1777         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1778             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1779                 struct inet_connection_sock *icsk = inet_csk(sk);
1780                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1781
1782                 tcp_enter_loss(sk, 1);
1783                 icsk->icsk_retransmits++;
1784                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1785                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1786                                           icsk->icsk_rto, TCP_RTO_MAX);
1787                 return 1;
1788         }
1789         return 0;
1790 }
1791
1792 static inline int tcp_fackets_out(struct tcp_sock *tp)
1793 {
1794         return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1795 }
1796
1797 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1798 {
1799         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1800 }
1801
1802 static inline int tcp_head_timedout(struct sock *sk)
1803 {
1804         struct tcp_sock *tp = tcp_sk(sk);
1805
1806         return tp->packets_out &&
1807                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1808 }
1809
1810 /* Linux NewReno/SACK/FACK/ECN state machine.
1811  * --------------------------------------
1812  *
1813  * "Open"       Normal state, no dubious events, fast path.
1814  * "Disorder"   In all the respects it is "Open",
1815  *              but requires a bit more attention. It is entered when
1816  *              we see some SACKs or dupacks. It is split of "Open"
1817  *              mainly to move some processing from fast path to slow one.
1818  * "CWR"        CWND was reduced due to some Congestion Notification event.
1819  *              It can be ECN, ICMP source quench, local device congestion.
1820  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1821  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1822  *
1823  * tcp_fastretrans_alert() is entered:
1824  * - each incoming ACK, if state is not "Open"
1825  * - when arrived ACK is unusual, namely:
1826  *      * SACK
1827  *      * Duplicate ACK.
1828  *      * ECN ECE.
1829  *
1830  * Counting packets in flight is pretty simple.
1831  *
1832  *      in_flight = packets_out - left_out + retrans_out
1833  *
1834  *      packets_out is SND.NXT-SND.UNA counted in packets.
1835  *
1836  *      retrans_out is number of retransmitted segments.
1837  *
1838  *      left_out is number of segments left network, but not ACKed yet.
1839  *
1840  *              left_out = sacked_out + lost_out
1841  *
1842  *     sacked_out: Packets, which arrived to receiver out of order
1843  *                 and hence not ACKed. With SACKs this number is simply
1844  *                 amount of SACKed data. Even without SACKs
1845  *                 it is easy to give pretty reliable estimate of this number,
1846  *                 counting duplicate ACKs.
1847  *
1848  *       lost_out: Packets lost by network. TCP has no explicit
1849  *                 "loss notification" feedback from network (for now).
1850  *                 It means that this number can be only _guessed_.
1851  *                 Actually, it is the heuristics to predict lossage that
1852  *                 distinguishes different algorithms.
1853  *
1854  *      F.e. after RTO, when all the queue is considered as lost,
1855  *      lost_out = packets_out and in_flight = retrans_out.
1856  *
1857  *              Essentially, we have now two algorithms counting
1858  *              lost packets.
1859  *
1860  *              FACK: It is the simplest heuristics. As soon as we decided
1861  *              that something is lost, we decide that _all_ not SACKed
1862  *              packets until the most forward SACK are lost. I.e.
1863  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1864  *              It is absolutely correct estimate, if network does not reorder
1865  *              packets. And it loses any connection to reality when reordering
1866  *              takes place. We use FACK by default until reordering
1867  *              is suspected on the path to this destination.
1868  *
1869  *              NewReno: when Recovery is entered, we assume that one segment
1870  *              is lost (classic Reno). While we are in Recovery and
1871  *              a partial ACK arrives, we assume that one more packet
1872  *              is lost (NewReno). This heuristics are the same in NewReno
1873  *              and SACK.
1874  *
1875  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1876  *  deflation etc. CWND is real congestion window, never inflated, changes
1877  *  only according to classic VJ rules.
1878  *
1879  * Really tricky (and requiring careful tuning) part of algorithm
1880  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1881  * The first determines the moment _when_ we should reduce CWND and,
1882  * hence, slow down forward transmission. In fact, it determines the moment
1883  * when we decide that hole is caused by loss, rather than by a reorder.
1884  *
1885  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1886  * holes, caused by lost packets.
1887  *
1888  * And the most logically complicated part of algorithm is undo
1889  * heuristics. We detect false retransmits due to both too early
1890  * fast retransmit (reordering) and underestimated RTO, analyzing
1891  * timestamps and D-SACKs. When we detect that some segments were
1892  * retransmitted by mistake and CWND reduction was wrong, we undo
1893  * window reduction and abort recovery phase. This logic is hidden
1894  * inside several functions named tcp_try_undo_<something>.
1895  */
1896
1897 /* This function decides, when we should leave Disordered state
1898  * and enter Recovery phase, reducing congestion window.
1899  *
1900  * Main question: may we further continue forward transmission
1901  * with the same cwnd?
1902  */
1903 static int tcp_time_to_recover(struct sock *sk)
1904 {
1905         struct tcp_sock *tp = tcp_sk(sk);
1906         __u32 packets_out;
1907
1908         /* Do not perform any recovery during FRTO algorithm */
1909         if (tp->frto_counter)
1910                 return 0;
1911
1912         /* Trick#1: The loss is proven. */
1913         if (tp->lost_out)
1914                 return 1;
1915
1916         /* Not-A-Trick#2 : Classic rule... */
1917         if (tcp_fackets_out(tp) > tp->reordering)
1918                 return 1;
1919
1920         /* Trick#3 : when we use RFC2988 timer restart, fast
1921          * retransmit can be triggered by timeout of queue head.
1922          */
1923         if (tcp_head_timedout(sk))
1924                 return 1;
1925
1926         /* Trick#4: It is still not OK... But will it be useful to delay
1927          * recovery more?
1928          */
1929         packets_out = tp->packets_out;
1930         if (packets_out <= tp->reordering &&
1931             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1932             !tcp_may_send_now(sk)) {
1933                 /* We have nothing to send. This connection is limited
1934                  * either by receiver window or by application.
1935                  */
1936                 return 1;
1937         }
1938
1939         return 0;
1940 }
1941
1942 /* RFC: This is from the original, I doubt that this is necessary at all:
1943  * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1944  * retransmitted past LOST markings in the first place? I'm not fully sure
1945  * about undo and end of connection cases, which can cause R without L?
1946  */
1947 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
1948                                        struct sk_buff *skb)
1949 {
1950         if ((tp->retransmit_skb_hint != NULL) &&
1951             before(TCP_SKB_CB(skb)->seq,
1952             TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1953                 tp->retransmit_skb_hint = NULL;
1954 }
1955
1956 /* Mark head of queue up as lost. */
1957 static void tcp_mark_head_lost(struct sock *sk,
1958                                int packets, u32 high_seq)
1959 {
1960         struct tcp_sock *tp = tcp_sk(sk);
1961         struct sk_buff *skb;
1962         int cnt;
1963
1964         BUG_TRAP(packets <= tp->packets_out);
1965         if (tp->lost_skb_hint) {
1966                 skb = tp->lost_skb_hint;
1967                 cnt = tp->lost_cnt_hint;
1968         } else {
1969                 skb = tcp_write_queue_head(sk);
1970                 cnt = 0;
1971         }
1972
1973         tcp_for_write_queue_from(skb, sk) {
1974                 if (skb == tcp_send_head(sk))
1975                         break;
1976                 /* TODO: do this better */
1977                 /* this is not the most efficient way to do this... */
1978                 tp->lost_skb_hint = skb;
1979                 tp->lost_cnt_hint = cnt;
1980                 cnt += tcp_skb_pcount(skb);
1981                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1982                         break;
1983                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1984                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1985                         tp->lost_out += tcp_skb_pcount(skb);
1986                         tcp_verify_retransmit_hint(tp, skb);
1987                 }
1988         }
1989         tcp_verify_left_out(tp);
1990 }
1991
1992 /* Account newly detected lost packet(s) */
1993
1994 static void tcp_update_scoreboard(struct sock *sk)
1995 {
1996         struct tcp_sock *tp = tcp_sk(sk);
1997
1998         if (tcp_is_fack(tp)) {
1999                 int lost = tp->fackets_out - tp->reordering;
2000                 if (lost <= 0)
2001                         lost = 1;
2002                 tcp_mark_head_lost(sk, lost, tp->high_seq);
2003         } else {
2004                 tcp_mark_head_lost(sk, 1, tp->high_seq);
2005         }
2006
2007         /* New heuristics: it is possible only after we switched
2008          * to restart timer each time when something is ACKed.
2009          * Hence, we can detect timed out packets during fast
2010          * retransmit without falling to slow start.
2011          */
2012         if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) {
2013                 struct sk_buff *skb;
2014
2015                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2016                         : tcp_write_queue_head(sk);
2017
2018                 tcp_for_write_queue_from(skb, sk) {
2019                         if (skb == tcp_send_head(sk))
2020                                 break;
2021                         if (!tcp_skb_timedout(sk, skb))
2022                                 break;
2023
2024                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
2025                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2026                                 tp->lost_out += tcp_skb_pcount(skb);
2027                                 tcp_verify_retransmit_hint(tp, skb);
2028                         }
2029                 }
2030
2031                 tp->scoreboard_skb_hint = skb;
2032
2033                 tcp_verify_left_out(tp);
2034         }
2035 }
2036
2037 /* CWND moderation, preventing bursts due to too big ACKs
2038  * in dubious situations.
2039  */
2040 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2041 {
2042         tp->snd_cwnd = min(tp->snd_cwnd,
2043                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
2044         tp->snd_cwnd_stamp = tcp_time_stamp;
2045 }
2046
2047 /* Lower bound on congestion window is slow start threshold
2048  * unless congestion avoidance choice decides to overide it.
2049  */
2050 static inline u32 tcp_cwnd_min(const struct sock *sk)
2051 {
2052         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2053
2054         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2055 }
2056
2057 /* Decrease cwnd each second ack. */
2058 static void tcp_cwnd_down(struct sock *sk, int flag)
2059 {
2060         struct tcp_sock *tp = tcp_sk(sk);
2061         int decr = tp->snd_cwnd_cnt + 1;
2062
2063         if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
2064             (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
2065                 tp->snd_cwnd_cnt = decr&1;
2066                 decr >>= 1;
2067
2068                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2069                         tp->snd_cwnd -= decr;
2070
2071                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
2072                 tp->snd_cwnd_stamp = tcp_time_stamp;
2073         }
2074 }
2075
2076 /* Nothing was retransmitted or returned timestamp is less
2077  * than timestamp of the first retransmission.
2078  */
2079 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2080 {
2081         return !tp->retrans_stamp ||
2082                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2083                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2084 }
2085
2086 /* Undo procedures. */
2087
2088 #if FASTRETRANS_DEBUG > 1
2089 static void DBGUNDO(struct sock *sk, const char *msg)
2090 {
2091         struct tcp_sock *tp = tcp_sk(sk);
2092         struct inet_sock *inet = inet_sk(sk);
2093
2094         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2095                msg,
2096                NIPQUAD(inet->daddr), ntohs(inet->dport),
2097                tp->snd_cwnd, tcp_left_out(tp),
2098                tp->snd_ssthresh, tp->prior_ssthresh,
2099                tp->packets_out);
2100 }
2101 #else
2102 #define DBGUNDO(x...) do { } while (0)
2103 #endif
2104
2105 static void tcp_undo_cwr(struct sock *sk, const int undo)
2106 {
2107         struct tcp_sock *tp = tcp_sk(sk);
2108
2109         if (tp->prior_ssthresh) {
2110                 const struct inet_connection_sock *icsk = inet_csk(sk);
2111
2112                 if (icsk->icsk_ca_ops->undo_cwnd)
2113                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2114                 else
2115                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
2116
2117                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2118                         tp->snd_ssthresh = tp->prior_ssthresh;
2119                         TCP_ECN_withdraw_cwr(tp);
2120                 }
2121         } else {
2122                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2123         }
2124         tcp_moderate_cwnd(tp);
2125         tp->snd_cwnd_stamp = tcp_time_stamp;
2126
2127         /* There is something screwy going on with the retrans hints after
2128            an undo */
2129         tcp_clear_all_retrans_hints(tp);
2130 }
2131
2132 static inline int tcp_may_undo(struct tcp_sock *tp)
2133 {
2134         return tp->undo_marker &&
2135                 (!tp->undo_retrans || tcp_packet_delayed(tp));
2136 }
2137
2138 /* People celebrate: "We love our President!" */
2139 static int tcp_try_undo_recovery(struct sock *sk)
2140 {
2141         struct tcp_sock *tp = tcp_sk(sk);
2142
2143         if (tcp_may_undo(tp)) {
2144                 /* Happy end! We did not retransmit anything
2145                  * or our original transmission succeeded.
2146                  */
2147                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2148                 tcp_undo_cwr(sk, 1);
2149                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2150                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2151                 else
2152                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2153                 tp->undo_marker = 0;
2154         }
2155         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2156                 /* Hold old state until something *above* high_seq
2157                  * is ACKed. For Reno it is MUST to prevent false
2158                  * fast retransmits (RFC2582). SACK TCP is safe. */
2159                 tcp_moderate_cwnd(tp);
2160                 return 1;
2161         }
2162         tcp_set_ca_state(sk, TCP_CA_Open);
2163         return 0;
2164 }
2165
2166 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2167 static void tcp_try_undo_dsack(struct sock *sk)
2168 {
2169         struct tcp_sock *tp = tcp_sk(sk);
2170
2171         if (tp->undo_marker && !tp->undo_retrans) {
2172                 DBGUNDO(sk, "D-SACK");
2173                 tcp_undo_cwr(sk, 1);
2174                 tp->undo_marker = 0;
2175                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2176         }
2177 }
2178
2179 /* Undo during fast recovery after partial ACK. */
2180
2181 static int tcp_try_undo_partial(struct sock *sk, int acked)
2182 {
2183         struct tcp_sock *tp = tcp_sk(sk);
2184         /* Partial ACK arrived. Force Hoe's retransmit. */
2185         int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering;
2186
2187         if (tcp_may_undo(tp)) {
2188                 /* Plain luck! Hole if filled with delayed
2189                  * packet, rather than with a retransmit.
2190                  */
2191                 if (tp->retrans_out == 0)
2192                         tp->retrans_stamp = 0;
2193
2194                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2195
2196                 DBGUNDO(sk, "Hoe");
2197                 tcp_undo_cwr(sk, 0);
2198                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2199
2200                 /* So... Do not make Hoe's retransmit yet.
2201                  * If the first packet was delayed, the rest
2202                  * ones are most probably delayed as well.
2203                  */
2204                 failed = 0;
2205         }
2206         return failed;
2207 }
2208
2209 /* Undo during loss recovery after partial ACK. */
2210 static int tcp_try_undo_loss(struct sock *sk)
2211 {
2212         struct tcp_sock *tp = tcp_sk(sk);
2213
2214         if (tcp_may_undo(tp)) {
2215                 struct sk_buff *skb;
2216                 tcp_for_write_queue(skb, sk) {
2217                         if (skb == tcp_send_head(sk))
2218                                 break;
2219                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2220                 }
2221
2222                 tcp_clear_all_retrans_hints(tp);
2223
2224                 DBGUNDO(sk, "partial loss");
2225                 tp->lost_out = 0;
2226                 tcp_undo_cwr(sk, 1);
2227                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2228                 inet_csk(sk)->icsk_retransmits = 0;
2229                 tp->undo_marker = 0;
2230                 if (tcp_is_sack(tp))
2231                         tcp_set_ca_state(sk, TCP_CA_Open);
2232                 return 1;
2233         }
2234         return 0;
2235 }
2236
2237 static inline void tcp_complete_cwr(struct sock *sk)
2238 {
2239         struct tcp_sock *tp = tcp_sk(sk);
2240         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2241         tp->snd_cwnd_stamp = tcp_time_stamp;
2242         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2243 }
2244
2245 static void tcp_try_to_open(struct sock *sk, int flag)
2246 {
2247         struct tcp_sock *tp = tcp_sk(sk);
2248
2249         tcp_verify_left_out(tp);
2250
2251         if (tp->retrans_out == 0)
2252                 tp->retrans_stamp = 0;
2253
2254         if (flag&FLAG_ECE)
2255                 tcp_enter_cwr(sk, 1);
2256
2257         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2258                 int state = TCP_CA_Open;
2259
2260                 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2261                         state = TCP_CA_Disorder;
2262
2263                 if (inet_csk(sk)->icsk_ca_state != state) {
2264                         tcp_set_ca_state(sk, state);
2265                         tp->high_seq = tp->snd_nxt;
2266                 }
2267                 tcp_moderate_cwnd(tp);
2268         } else {
2269                 tcp_cwnd_down(sk, flag);
2270         }
2271 }
2272
2273 static void tcp_mtup_probe_failed(struct sock *sk)
2274 {
2275         struct inet_connection_sock *icsk = inet_csk(sk);
2276
2277         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2278         icsk->icsk_mtup.probe_size = 0;
2279 }
2280
2281 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2282 {
2283         struct tcp_sock *tp = tcp_sk(sk);
2284         struct inet_connection_sock *icsk = inet_csk(sk);
2285
2286         /* FIXME: breaks with very large cwnd */
2287         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2288         tp->snd_cwnd = tp->snd_cwnd *
2289                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2290                        icsk->icsk_mtup.probe_size;
2291         tp->snd_cwnd_cnt = 0;
2292         tp->snd_cwnd_stamp = tcp_time_stamp;
2293         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2294
2295         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2296         icsk->icsk_mtup.probe_size = 0;
2297         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2298 }
2299
2300
2301 /* Process an event, which can update packets-in-flight not trivially.
2302  * Main goal of this function is to calculate new estimate for left_out,
2303  * taking into account both packets sitting in receiver's buffer and
2304  * packets lost by network.
2305  *
2306  * Besides that it does CWND reduction, when packet loss is detected
2307  * and changes state of machine.
2308  *
2309  * It does _not_ decide what to send, it is made in function
2310  * tcp_xmit_retransmit_queue().
2311  */
2312 static void
2313 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2314 {
2315         struct inet_connection_sock *icsk = inet_csk(sk);
2316         struct tcp_sock *tp = tcp_sk(sk);
2317         int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2318         int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2319                                     (tp->fackets_out > tp->reordering));
2320
2321         /* Some technical things:
2322          * 1. Reno does not count dupacks (sacked_out) automatically. */
2323         if (!tp->packets_out)
2324                 tp->sacked_out = 0;
2325
2326         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2327                 tp->fackets_out = 0;
2328
2329         /* Now state machine starts.
2330          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2331         if (flag&FLAG_ECE)
2332                 tp->prior_ssthresh = 0;
2333
2334         /* B. In all the states check for reneging SACKs. */
2335         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2336                 return;
2337
2338         /* C. Process data loss notification, provided it is valid. */
2339         if ((flag&FLAG_DATA_LOST) &&
2340             before(tp->snd_una, tp->high_seq) &&
2341             icsk->icsk_ca_state != TCP_CA_Open &&
2342             tp->fackets_out > tp->reordering) {
2343                 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2344                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2345         }
2346
2347         /* D. Check consistency of the current state. */
2348         tcp_verify_left_out(tp);
2349
2350         /* E. Check state exit conditions. State can be terminated
2351          *    when high_seq is ACKed. */
2352         if (icsk->icsk_ca_state == TCP_CA_Open) {
2353                 BUG_TRAP(tp->retrans_out == 0);
2354                 tp->retrans_stamp = 0;
2355         } else if (!before(tp->snd_una, tp->high_seq)) {
2356                 switch (icsk->icsk_ca_state) {
2357                 case TCP_CA_Loss:
2358                         icsk->icsk_retransmits = 0;
2359                         if (tcp_try_undo_recovery(sk))
2360                                 return;
2361                         break;
2362
2363                 case TCP_CA_CWR:
2364                         /* CWR is to be held something *above* high_seq
2365                          * is ACKed for CWR bit to reach receiver. */
2366                         if (tp->snd_una != tp->high_seq) {
2367                                 tcp_complete_cwr(sk);
2368                                 tcp_set_ca_state(sk, TCP_CA_Open);
2369                         }
2370                         break;
2371
2372                 case TCP_CA_Disorder:
2373                         tcp_try_undo_dsack(sk);
2374                         if (!tp->undo_marker ||
2375                             /* For SACK case do not Open to allow to undo
2376                              * catching for all duplicate ACKs. */
2377                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2378                                 tp->undo_marker = 0;
2379                                 tcp_set_ca_state(sk, TCP_CA_Open);
2380                         }
2381                         break;
2382
2383                 case TCP_CA_Recovery:
2384                         if (tcp_is_reno(tp))
2385                                 tcp_reset_reno_sack(tp);
2386                         if (tcp_try_undo_recovery(sk))
2387                                 return;
2388                         tcp_complete_cwr(sk);
2389                         break;
2390                 }
2391         }
2392
2393         /* F. Process state. */
2394         switch (icsk->icsk_ca_state) {
2395         case TCP_CA_Recovery:
2396                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2397                         if (tcp_is_reno(tp) && is_dupack)
2398                                 tcp_add_reno_sack(sk);
2399                 } else
2400                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2401                 break;
2402         case TCP_CA_Loss:
2403                 if (flag&FLAG_DATA_ACKED)
2404                         icsk->icsk_retransmits = 0;
2405                 if (!tcp_try_undo_loss(sk)) {
2406                         tcp_moderate_cwnd(tp);
2407                         tcp_xmit_retransmit_queue(sk);
2408                         return;
2409                 }
2410                 if (icsk->icsk_ca_state != TCP_CA_Open)
2411                         return;
2412                 /* Loss is undone; fall through to processing in Open state. */
2413         default:
2414                 if (tcp_is_reno(tp)) {
2415                         if (flag & FLAG_SND_UNA_ADVANCED)
2416                                 tcp_reset_reno_sack(tp);
2417                         if (is_dupack)
2418                                 tcp_add_reno_sack(sk);
2419                 }
2420
2421                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2422                         tcp_try_undo_dsack(sk);
2423
2424                 if (!tcp_time_to_recover(sk)) {
2425                         tcp_try_to_open(sk, flag);
2426                         return;
2427                 }
2428
2429                 /* MTU probe failure: don't reduce cwnd */
2430                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2431                     icsk->icsk_mtup.probe_size &&
2432                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2433                         tcp_mtup_probe_failed(sk);
2434                         /* Restores the reduction we did in tcp_mtup_probe() */
2435                         tp->snd_cwnd++;
2436                         tcp_simple_retransmit(sk);
2437                         return;
2438                 }
2439
2440                 /* Otherwise enter Recovery state */
2441
2442                 if (tcp_is_reno(tp))
2443                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2444                 else
2445                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2446
2447                 tp->high_seq = tp->snd_nxt;
2448                 tp->prior_ssthresh = 0;
2449                 tp->undo_marker = tp->snd_una;
2450                 tp->undo_retrans = tp->retrans_out;
2451
2452                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2453                         if (!(flag&FLAG_ECE))
2454                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2455                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2456                         TCP_ECN_queue_cwr(tp);
2457                 }
2458
2459                 tp->bytes_acked = 0;
2460                 tp->snd_cwnd_cnt = 0;
2461                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2462         }
2463
2464         if (do_lost || tcp_head_timedout(sk))
2465                 tcp_update_scoreboard(sk);
2466         tcp_cwnd_down(sk, flag);
2467         tcp_xmit_retransmit_queue(sk);
2468 }
2469
2470 /* Read draft-ietf-tcplw-high-performance before mucking
2471  * with this code. (Supersedes RFC1323)
2472  */
2473 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2474 {
2475         /* RTTM Rule: A TSecr value received in a segment is used to
2476          * update the averaged RTT measurement only if the segment
2477          * acknowledges some new data, i.e., only if it advances the
2478          * left edge of the send window.
2479          *
2480          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2481          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2482          *
2483          * Changed: reset backoff as soon as we see the first valid sample.
2484          * If we do not, we get strongly overestimated rto. With timestamps
2485          * samples are accepted even from very old segments: f.e., when rtt=1
2486          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2487          * answer arrives rto becomes 120 seconds! If at least one of segments
2488          * in window is lost... Voila.                          --ANK (010210)
2489          */
2490         struct tcp_sock *tp = tcp_sk(sk);
2491         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2492         tcp_rtt_estimator(sk, seq_rtt);
2493         tcp_set_rto(sk);
2494         inet_csk(sk)->icsk_backoff = 0;
2495         tcp_bound_rto(sk);
2496 }
2497
2498 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2499 {
2500         /* We don't have a timestamp. Can only use
2501          * packets that are not retransmitted to determine
2502          * rtt estimates. Also, we must not reset the
2503          * backoff for rto until we get a non-retransmitted
2504          * packet. This allows us to deal with a situation
2505          * where the network delay has increased suddenly.
2506          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2507          */
2508
2509         if (flag & FLAG_RETRANS_DATA_ACKED)
2510                 return;
2511
2512         tcp_rtt_estimator(sk, seq_rtt);
2513         tcp_set_rto(sk);
2514         inet_csk(sk)->icsk_backoff = 0;
2515         tcp_bound_rto(sk);
2516 }
2517
2518 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2519                                       const s32 seq_rtt)
2520 {
2521         const struct tcp_sock *tp = tcp_sk(sk);
2522         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2523         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2524                 tcp_ack_saw_tstamp(sk, flag);
2525         else if (seq_rtt >= 0)
2526                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2527 }
2528
2529 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2530                            u32 in_flight, int good)
2531 {
2532         const struct inet_connection_sock *icsk = inet_csk(sk);
2533         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2534         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2535 }
2536
2537 /* Restart timer after forward progress on connection.
2538  * RFC2988 recommends to restart timer to now+rto.
2539  */
2540 static void tcp_rearm_rto(struct sock *sk)
2541 {
2542         struct tcp_sock *tp = tcp_sk(sk);
2543
2544         if (!tp->packets_out) {
2545                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2546         } else {
2547                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2548         }
2549 }
2550
2551 /* If we get here, the whole TSO packet has not been acked. */
2552 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2553 {
2554         struct tcp_sock *tp = tcp_sk(sk);
2555         u32 packets_acked;
2556
2557         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2558
2559         packets_acked = tcp_skb_pcount(skb);
2560         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2561                 return 0;
2562         packets_acked -= tcp_skb_pcount(skb);
2563
2564         if (packets_acked) {
2565                 BUG_ON(tcp_skb_pcount(skb) == 0);
2566                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2567         }
2568
2569         return packets_acked;
2570 }
2571
2572 /* Remove acknowledged frames from the retransmission queue. If our packet
2573  * is before the ack sequence we can discard it as it's confirmed to have
2574  * arrived at the other end.
2575  */
2576 static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p)
2577 {
2578         struct tcp_sock *tp = tcp_sk(sk);
2579         const struct inet_connection_sock *icsk = inet_csk(sk);
2580         struct sk_buff *skb;
2581         u32 now = tcp_time_stamp;
2582         int fully_acked = 1;
2583         int flag = 0;
2584         int prior_packets = tp->packets_out;
2585         s32 seq_rtt = -1;
2586         ktime_t last_ackt = net_invalid_timestamp();
2587
2588         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2589                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2590                 u32 end_seq;
2591                 u32 packets_acked;
2592                 u8 sacked = scb->sacked;
2593
2594                 if (after(scb->end_seq, tp->snd_una)) {
2595                         if (tcp_skb_pcount(skb) == 1 ||
2596                             !after(tp->snd_una, scb->seq))
2597                                 break;
2598
2599                         packets_acked = tcp_tso_acked(sk, skb);
2600                         if (!packets_acked)
2601                                 break;
2602
2603                         fully_acked = 0;
2604                         end_seq = tp->snd_una;
2605                 } else {
2606                         packets_acked = tcp_skb_pcount(skb);
2607                         end_seq = scb->end_seq;
2608                 }
2609
2610                 /* MTU probing checks */
2611                 if (fully_acked && icsk->icsk_mtup.probe_size &&
2612                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2613                         tcp_mtup_probe_success(sk, skb);
2614                 }
2615
2616                 if (sacked) {
2617                         if (sacked & TCPCB_RETRANS) {
2618                                 if (sacked & TCPCB_SACKED_RETRANS)
2619                                         tp->retrans_out -= packets_acked;
2620                                 flag |= FLAG_RETRANS_DATA_ACKED;
2621                                 seq_rtt = -1;
2622                                 if ((flag & FLAG_DATA_ACKED) ||
2623                                     (packets_acked > 1))
2624                                         flag |= FLAG_NONHEAD_RETRANS_ACKED;
2625                         } else if (seq_rtt < 0) {
2626                                 seq_rtt = now - scb->when;
2627                                 if (fully_acked)
2628                                         last_ackt = skb->tstamp;
2629                         }
2630
2631                         if (sacked & TCPCB_SACKED_ACKED)
2632                                 tp->sacked_out -= packets_acked;
2633                         if (sacked & TCPCB_LOST)
2634                                 tp->lost_out -= packets_acked;
2635
2636                         if ((sacked & TCPCB_URG) && tp->urg_mode &&
2637                             !before(end_seq, tp->snd_up))
2638                                 tp->urg_mode = 0;
2639                 } else if (seq_rtt < 0) {
2640                         seq_rtt = now - scb->when;
2641                         if (fully_acked)
2642                                 last_ackt = skb->tstamp;
2643                 }
2644                 tp->packets_out -= packets_acked;
2645
2646                 /* Initial outgoing SYN's get put onto the write_queue
2647                  * just like anything else we transmit.  It is not
2648                  * true data, and if we misinform our callers that
2649                  * this ACK acks real data, we will erroneously exit
2650                  * connection startup slow start one packet too
2651                  * quickly.  This is severely frowned upon behavior.
2652                  */
2653                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2654                         flag |= FLAG_DATA_ACKED;
2655                 } else {
2656                         flag |= FLAG_SYN_ACKED;
2657                         tp->retrans_stamp = 0;
2658                 }
2659
2660                 if (!fully_acked)
2661                         break;
2662
2663                 tcp_unlink_write_queue(skb, sk);
2664                 sk_stream_free_skb(sk, skb);
2665                 tcp_clear_all_retrans_hints(tp);
2666         }
2667
2668         if (flag & FLAG_ACKED) {
2669                 u32 pkts_acked = prior_packets - tp->packets_out;
2670                 const struct tcp_congestion_ops *ca_ops
2671                         = inet_csk(sk)->icsk_ca_ops;
2672
2673                 tcp_ack_update_rtt(sk, flag, seq_rtt);
2674                 tcp_rearm_rto(sk);
2675
2676                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2677                 /* hint's skb might be NULL but we don't need to care */
2678                 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
2679                                                tp->fastpath_cnt_hint);
2680                 if (tcp_is_reno(tp))
2681                         tcp_remove_reno_sacks(sk, pkts_acked);
2682
2683                 if (ca_ops->pkts_acked) {
2684                         s32 rtt_us = -1;
2685
2686                         /* Is the ACK triggering packet unambiguous? */
2687                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2688                                 /* High resolution needed and available? */
2689                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2690                                     !ktime_equal(last_ackt,
2691                                                  net_invalid_timestamp()))
2692                                         rtt_us = ktime_us_delta(ktime_get_real(),
2693                                                                 last_ackt);
2694                                 else if (seq_rtt > 0)
2695                                         rtt_us = jiffies_to_usecs(seq_rtt);
2696                         }
2697
2698                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2699                 }
2700         }
2701
2702 #if FASTRETRANS_DEBUG > 0
2703         BUG_TRAP((int)tp->sacked_out >= 0);
2704         BUG_TRAP((int)tp->lost_out >= 0);
2705         BUG_TRAP((int)tp->retrans_out >= 0);
2706         if (!tp->packets_out && tcp_is_sack(tp)) {
2707                 icsk = inet_csk(sk);
2708                 if (tp->lost_out) {
2709                         printk(KERN_DEBUG "Leak l=%u %d\n",
2710                                tp->lost_out, icsk->icsk_ca_state);
2711                         tp->lost_out = 0;
2712                 }
2713                 if (tp->sacked_out) {
2714                         printk(KERN_DEBUG "Leak s=%u %d\n",
2715                                tp->sacked_out, icsk->icsk_ca_state);
2716                         tp->sacked_out = 0;
2717                 }
2718                 if (tp->retrans_out) {
2719                         printk(KERN_DEBUG "Leak r=%u %d\n",
2720                                tp->retrans_out, icsk->icsk_ca_state);
2721                         tp->retrans_out = 0;
2722                 }
2723         }
2724 #endif
2725         *seq_rtt_p = seq_rtt;
2726         return flag;
2727 }
2728
2729 static void tcp_ack_probe(struct sock *sk)
2730 {
2731         const struct tcp_sock *tp = tcp_sk(sk);
2732         struct inet_connection_sock *icsk = inet_csk(sk);
2733
2734         /* Was it a usable window open? */
2735
2736         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2737                    tp->snd_una + tp->snd_wnd)) {
2738                 icsk->icsk_backoff = 0;
2739                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2740                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2741                  * This function is not for random using!
2742                  */
2743         } else {
2744                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2745                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2746                                           TCP_RTO_MAX);
2747         }
2748 }
2749
2750 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2751 {
2752         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2753                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2754 }
2755
2756 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2757 {
2758         const struct tcp_sock *tp = tcp_sk(sk);
2759         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2760                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2761 }
2762
2763 /* Check that window update is acceptable.
2764  * The function assumes that snd_una<=ack<=snd_next.
2765  */
2766 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2767                                         const u32 ack_seq, const u32 nwin)
2768 {
2769         return (after(ack, tp->snd_una) ||
2770                 after(ack_seq, tp->snd_wl1) ||
2771                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2772 }
2773
2774 /* Update our send window.
2775  *
2776  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2777  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2778  */
2779 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2780                                  u32 ack_seq)
2781 {
2782         struct tcp_sock *tp = tcp_sk(sk);
2783         int flag = 0;
2784         u32 nwin = ntohs(tcp_hdr(skb)->window);
2785
2786         if (likely(!tcp_hdr(skb)->syn))
2787                 nwin <<= tp->rx_opt.snd_wscale;
2788
2789         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2790                 flag |= FLAG_WIN_UPDATE;
2791                 tcp_update_wl(tp, ack, ack_seq);
2792
2793                 if (tp->snd_wnd != nwin) {
2794                         tp->snd_wnd = nwin;
2795
2796                         /* Note, it is the only place, where
2797                          * fast path is recovered for sending TCP.
2798                          */
2799                         tp->pred_flags = 0;
2800                         tcp_fast_path_check(sk);
2801
2802                         if (nwin > tp->max_window) {
2803                                 tp->max_window = nwin;
2804                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2805                         }
2806                 }
2807         }
2808
2809         tp->snd_una = ack;
2810
2811         return flag;
2812 }
2813
2814 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2815  * continue in congestion avoidance.
2816  */
2817 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2818 {
2819         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2820         tp->snd_cwnd_cnt = 0;
2821         TCP_ECN_queue_cwr(tp);
2822         tcp_moderate_cwnd(tp);
2823 }
2824
2825 /* A conservative spurious RTO response algorithm: reduce cwnd using
2826  * rate halving and continue in congestion avoidance.
2827  */
2828 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2829 {
2830         tcp_enter_cwr(sk, 0);
2831 }
2832
2833 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2834 {
2835         if (flag&FLAG_ECE)
2836                 tcp_ratehalving_spur_to_response(sk);
2837         else
2838                 tcp_undo_cwr(sk, 1);
2839 }
2840
2841 /* F-RTO spurious RTO detection algorithm (RFC4138)
2842  *
2843  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2844  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2845  * window (but not to or beyond highest sequence sent before RTO):
2846  *   On First ACK,  send two new segments out.
2847  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2848  *                  algorithm is not part of the F-RTO detection algorithm
2849  *                  given in RFC4138 but can be selected separately).
2850  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2851  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2852  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2853  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2854  *
2855  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2856  * original window even after we transmit two new data segments.
2857  *
2858  * SACK version:
2859  *   on first step, wait until first cumulative ACK arrives, then move to
2860  *   the second step. In second step, the next ACK decides.
2861  *
2862  * F-RTO is implemented (mainly) in four functions:
2863  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2864  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2865  *     called when tcp_use_frto() showed green light
2866  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2867  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2868  *     to prove that the RTO is indeed spurious. It transfers the control
2869  *     from F-RTO to the conventional RTO recovery
2870  */
2871 static int tcp_process_frto(struct sock *sk, int flag)
2872 {
2873         struct tcp_sock *tp = tcp_sk(sk);
2874
2875         tcp_verify_left_out(tp);
2876
2877         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2878         if (flag&FLAG_DATA_ACKED)
2879                 inet_csk(sk)->icsk_retransmits = 0;
2880
2881         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
2882             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
2883                 tp->undo_marker = 0;
2884
2885         if (!before(tp->snd_una, tp->frto_highmark)) {
2886                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2887                 return 1;
2888         }
2889
2890         if (!IsSackFrto() || tcp_is_reno(tp)) {
2891                 /* RFC4138 shortcoming in step 2; should also have case c):
2892                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2893                  * data, winupdate
2894                  */
2895                 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2896                         return 1;
2897
2898                 if (!(flag&FLAG_DATA_ACKED)) {
2899                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2900                                             flag);
2901                         return 1;
2902                 }
2903         } else {
2904                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2905                         /* Prevent sending of new data. */
2906                         tp->snd_cwnd = min(tp->snd_cwnd,
2907                                            tcp_packets_in_flight(tp));
2908                         return 1;
2909                 }
2910
2911                 if ((tp->frto_counter >= 2) &&
2912                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2913                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2914                         /* RFC4138 shortcoming (see comment above) */
2915                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2916                                 return 1;
2917
2918                         tcp_enter_frto_loss(sk, 3, flag);
2919                         return 1;
2920                 }
2921         }
2922
2923         if (tp->frto_counter == 1) {
2924                 /* Sending of the next skb must be allowed or no FRTO */
2925                 if (!tcp_send_head(sk) ||
2926                     after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2927                                      tp->snd_una + tp->snd_wnd)) {
2928                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2929                                             flag);
2930                         return 1;
2931                 }
2932
2933                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2934                 tp->frto_counter = 2;
2935                 return 1;
2936         } else {
2937                 switch (sysctl_tcp_frto_response) {
2938                 case 2:
2939                         tcp_undo_spur_to_response(sk, flag);
2940                         break;
2941                 case 1:
2942                         tcp_conservative_spur_to_response(tp);
2943                         break;
2944                 default:
2945                         tcp_ratehalving_spur_to_response(sk);
2946                         break;
2947                 }
2948                 tp->frto_counter = 0;
2949                 tp->undo_marker = 0;
2950                 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
2951         }
2952         return 0;
2953 }
2954
2955 /* This routine deals with incoming acks, but not outgoing ones. */
2956 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2957 {
2958         struct inet_connection_sock *icsk = inet_csk(sk);
2959         struct tcp_sock *tp = tcp_sk(sk);
2960         u32 prior_snd_una = tp->snd_una;
2961         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2962         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2963         u32 prior_in_flight;
2964         s32 seq_rtt;
2965         int prior_packets;
2966         int frto_cwnd = 0;
2967
2968         /* If the ack is newer than sent or older than previous acks
2969          * then we can probably ignore it.
2970          */
2971         if (after(ack, tp->snd_nxt))
2972                 goto uninteresting_ack;
2973
2974         if (before(ack, prior_snd_una))
2975                 goto old_ack;
2976
2977         if (after(ack, prior_snd_una))
2978                 flag |= FLAG_SND_UNA_ADVANCED;
2979
2980         if (sysctl_tcp_abc) {
2981                 if (icsk->icsk_ca_state < TCP_CA_CWR)
2982                         tp->bytes_acked += ack - prior_snd_una;
2983                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2984                         /* we assume just one segment left network */
2985                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2986         }
2987
2988         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2989                 /* Window is constant, pure forward advance.
2990                  * No more checks are required.
2991                  * Note, we use the fact that SND.UNA>=SND.WL2.
2992                  */
2993                 tcp_update_wl(tp, ack, ack_seq);
2994                 tp->snd_una = ack;
2995                 flag |= FLAG_WIN_UPDATE;
2996
2997                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2998
2999                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3000         } else {
3001                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3002                         flag |= FLAG_DATA;
3003                 else
3004                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3005
3006                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3007
3008                 if (TCP_SKB_CB(skb)->sacked)
3009                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3010
3011                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3012                         flag |= FLAG_ECE;
3013
3014                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3015         }
3016
3017         /* We passed data and got it acked, remove any soft error
3018          * log. Something worked...
3019          */
3020         sk->sk_err_soft = 0;
3021         tp->rcv_tstamp = tcp_time_stamp;
3022         prior_packets = tp->packets_out;
3023         if (!prior_packets)
3024                 goto no_queue;
3025
3026         prior_in_flight = tcp_packets_in_flight(tp);
3027
3028         /* See if we can take anything off of the retransmit queue. */
3029         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
3030
3031         /* Guarantee sacktag reordering detection against wrap-arounds */
3032         if (before(tp->frto_highmark, tp->snd_una))
3033                 tp->frto_highmark = 0;
3034         if (tp->frto_counter)
3035                 frto_cwnd = tcp_process_frto(sk, flag);
3036
3037         if (tcp_ack_is_dubious(sk, flag)) {
3038                 /* Advance CWND, if state allows this. */
3039                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3040                     tcp_may_raise_cwnd(sk, flag))
3041                         tcp_cong_avoid(sk, ack, prior_in_flight, 0);
3042                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
3043         } else {
3044                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3045                         tcp_cong_avoid(sk, ack, prior_in_flight, 1);
3046         }
3047
3048         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
3049                 dst_confirm(sk->sk_dst_cache);
3050
3051         return 1;
3052
3053 no_queue:
3054         icsk->icsk_probes_out = 0;
3055
3056         /* If this ack opens up a zero window, clear backoff.  It was
3057          * being used to time the probes, and is probably far higher than
3058          * it needs to be for normal retransmission.
3059          */
3060         if (tcp_send_head(sk))
3061                 tcp_ack_probe(sk);
3062         return 1;
3063
3064 old_ack:
3065         if (TCP_SKB_CB(skb)->sacked)
3066                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3067
3068 uninteresting_ack:
3069         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3070         return 0;
3071 }
3072
3073
3074 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3075  * But, this can also be called on packets in the established flow when
3076  * the fast version below fails.
3077  */
3078 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
3079 {
3080         unsigned char *ptr;
3081         struct tcphdr *th = tcp_hdr(skb);
3082         int length=(th->doff*4)-sizeof(struct tcphdr);
3083
3084         ptr = (unsigned char *)(th + 1);
3085         opt_rx->saw_tstamp = 0;
3086
3087         while (length > 0) {
3088                 int opcode=*ptr++;
3089                 int opsize;
3090
3091                 switch (opcode) {
3092                         case TCPOPT_EOL:
3093                                 return;
3094                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3095                                 length--;
3096                                 continue;
3097                         default:
3098                                 opsize=*ptr++;
3099                                 if (opsize < 2) /* "silly options" */
3100                                         return;
3101                                 if (opsize > length)
3102                                         return; /* don't parse partial options */
3103                                 switch (opcode) {
3104                                 case TCPOPT_MSS:
3105                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
3106                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3107                                                 if (in_mss) {
3108                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3109                                                                 in_mss = opt_rx->user_mss;
3110                                                         opt_rx->mss_clamp = in_mss;
3111                                                 }
3112                                         }
3113                                         break;
3114                                 case TCPOPT_WINDOW:
3115                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3116                                                 if (sysctl_tcp_window_scaling) {
3117                                                         __u8 snd_wscale = *(__u8 *) ptr;
3118                                                         opt_rx->wscale_ok = 1;
3119                                                         if (snd_wscale > 14) {
3120                                                                 if (net_ratelimit())
3121                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3122                                                                                "scaling value %d >14 received.\n",
3123                                                                                snd_wscale);
3124                                                                 snd_wscale = 14;
3125                                                         }
3126                                                         opt_rx->snd_wscale = snd_wscale;
3127                                                 }
3128                                         break;
3129                                 case TCPOPT_TIMESTAMP:
3130                                         if (opsize==TCPOLEN_TIMESTAMP) {
3131                                                 if ((estab && opt_rx->tstamp_ok) ||
3132                                                     (!estab && sysctl_tcp_timestamps)) {
3133                                                         opt_rx->saw_tstamp = 1;
3134                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3135                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3136                                                 }
3137                                         }
3138                                         break;
3139                                 case TCPOPT_SACK_PERM:
3140                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3141                                                 if (sysctl_tcp_sack) {
3142                                                         opt_rx->sack_ok = 1;
3143                                                         tcp_sack_reset(opt_rx);
3144                                                 }
3145                                         }
3146                                         break;
3147
3148                                 case TCPOPT_SACK:
3149                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3150                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3151                                            opt_rx->sack_ok) {
3152                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3153                                         }
3154                                         break;
3155 #ifdef CONFIG_TCP_MD5SIG
3156                                 case TCPOPT_MD5SIG:
3157                                         /*
3158                                          * The MD5 Hash has already been
3159                                          * checked (see tcp_v{4,6}_do_rcv()).
3160                                          */
3161                                         break;
3162 #endif
3163                                 }
3164
3165                                 ptr+=opsize-2;
3166                                 length-=opsize;
3167                 }
3168         }
3169 }
3170
3171 /* Fast parse options. This hopes to only see timestamps.
3172  * If it is wrong it falls back on tcp_parse_options().
3173  */
3174 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3175                                   struct tcp_sock *tp)
3176 {
3177         if (th->doff == sizeof(struct tcphdr)>>2) {
3178                 tp->rx_opt.saw_tstamp = 0;
3179                 return 0;
3180         } else if (tp->rx_opt.tstamp_ok &&
3181                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3182                 __be32 *ptr = (__be32 *)(th + 1);
3183                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3184                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3185                         tp->rx_opt.saw_tstamp = 1;
3186                         ++ptr;
3187                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3188                         ++ptr;
3189                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3190                         return 1;
3191                 }
3192         }
3193         tcp_parse_options(skb, &tp->rx_opt, 1);
3194         return 1;
3195 }
3196
3197 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3198 {
3199         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3200         tp->rx_opt.ts_recent_stamp = get_seconds();
3201 }
3202
3203 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3204 {
3205         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3206                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3207                  * extra check below makes sure this can only happen
3208                  * for pure ACK frames.  -DaveM
3209                  *
3210                  * Not only, also it occurs for expired timestamps.
3211                  */
3212
3213                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3214                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3215                         tcp_store_ts_recent(tp);
3216         }
3217 }
3218
3219 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3220  *
3221  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3222  * it can pass through stack. So, the following predicate verifies that
3223  * this segment is not used for anything but congestion avoidance or
3224  * fast retransmit. Moreover, we even are able to eliminate most of such
3225  * second order effects, if we apply some small "replay" window (~RTO)
3226  * to timestamp space.
3227  *
3228  * All these measures still do not guarantee that we reject wrapped ACKs
3229  * on networks with high bandwidth, when sequence space is recycled fastly,
3230  * but it guarantees that such events will be very rare and do not affect
3231  * connection seriously. This doesn't look nice, but alas, PAWS is really
3232  * buggy extension.
3233  *
3234  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3235  * states that events when retransmit arrives after original data are rare.
3236  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3237  * the biggest problem on large power networks even with minor reordering.
3238  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3239  * up to bandwidth of 18Gigabit/sec. 8) ]
3240  */
3241
3242 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3243 {
3244         struct tcp_sock *tp = tcp_sk(sk);
3245         struct tcphdr *th = tcp_hdr(skb);
3246         u32 seq = TCP_SKB_CB(skb)->seq;
3247         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3248
3249         return (/* 1. Pure ACK with correct sequence number. */
3250                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3251
3252                 /* 2. ... and duplicate ACK. */
3253                 ack == tp->snd_una &&
3254
3255                 /* 3. ... and does not update window. */
3256                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3257
3258                 /* 4. ... and sits in replay window. */
3259                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3260 }
3261
3262 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3263 {
3264         const struct tcp_sock *tp = tcp_sk(sk);
3265         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3266                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3267                 !tcp_disordered_ack(sk, skb));
3268 }
3269
3270 /* Check segment sequence number for validity.
3271  *
3272  * Segment controls are considered valid, if the segment
3273  * fits to the window after truncation to the window. Acceptability
3274  * of data (and SYN, FIN, of course) is checked separately.
3275  * See tcp_data_queue(), for example.
3276  *
3277  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3278  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3279  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3280  * (borrowed from freebsd)
3281  */
3282
3283 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3284 {
3285         return  !before(end_seq, tp->rcv_wup) &&
3286                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3287 }
3288
3289 /* When we get a reset we do this. */
3290 static void tcp_reset(struct sock *sk)
3291 {
3292         /* We want the right error as BSD sees it (and indeed as we do). */
3293         switch (sk->sk_state) {
3294                 case TCP_SYN_SENT:
3295                         sk->sk_err = ECONNREFUSED;
3296                         break;
3297                 case TCP_CLOSE_WAIT:
3298                         sk->sk_err = EPIPE;
3299                         break;
3300                 case TCP_CLOSE:
3301                         return;
3302                 default:
3303                         sk->sk_err = ECONNRESET;
3304         }
3305
3306         if (!sock_flag(sk, SOCK_DEAD))
3307                 sk->sk_error_report(sk);
3308
3309         tcp_done(sk);
3310 }
3311
3312 /*
3313  *      Process the FIN bit. This now behaves as it is supposed to work
3314  *      and the FIN takes effect when it is validly part of sequence
3315  *      space. Not before when we get holes.
3316  *
3317  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3318  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3319  *      TIME-WAIT)
3320  *
3321  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3322  *      close and we go into CLOSING (and later onto TIME-WAIT)
3323  *
3324  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3325  */
3326 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3327 {
3328         struct tcp_sock *tp = tcp_sk(sk);
3329
3330         inet_csk_schedule_ack(sk);
3331
3332         sk->sk_shutdown |= RCV_SHUTDOWN;
3333         sock_set_flag(sk, SOCK_DONE);
3334
3335         switch (sk->sk_state) {
3336                 case TCP_SYN_RECV:
3337                 case TCP_ESTABLISHED:
3338                         /* Move to CLOSE_WAIT */
3339                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3340                         inet_csk(sk)->icsk_ack.pingpong = 1;
3341                         break;
3342
3343                 case TCP_CLOSE_WAIT:
3344                 case TCP_CLOSING:
3345                         /* Received a retransmission of the FIN, do
3346                          * nothing.
3347                          */
3348                         break;
3349                 case TCP_LAST_ACK:
3350                         /* RFC793: Remain in the LAST-ACK state. */
3351                         break;
3352
3353                 case TCP_FIN_WAIT1:
3354                         /* This case occurs when a simultaneous close
3355                          * happens, we must ack the received FIN and
3356                          * enter the CLOSING state.
3357                          */
3358                         tcp_send_ack(sk);
3359                         tcp_set_state(sk, TCP_CLOSING);
3360                         break;
3361                 case TCP_FIN_WAIT2:
3362                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3363                         tcp_send_ack(sk);
3364                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3365                         break;
3366                 default:
3367                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3368                          * cases we should never reach this piece of code.
3369                          */
3370                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3371                                __FUNCTION__, sk->sk_state);
3372                         break;
3373         }
3374
3375         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3376          * Probably, we should reset in this case. For now drop them.
3377          */
3378         __skb_queue_purge(&tp->out_of_order_queue);
3379         if (tcp_is_sack(tp))
3380                 tcp_sack_reset(&tp->rx_opt);
3381         sk_stream_mem_reclaim(sk);
3382
3383         if (!sock_flag(sk, SOCK_DEAD)) {
3384                 sk->sk_state_change(sk);
3385
3386                 /* Do not send POLL_HUP for half duplex close. */
3387                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3388                     sk->sk_state == TCP_CLOSE)
3389                         sk_wake_async(sk, 1, POLL_HUP);
3390                 else
3391                         sk_wake_async(sk, 1, POLL_IN);
3392         }
3393 }
3394
3395 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3396 {
3397         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3398                 if (before(seq, sp->start_seq))
3399                         sp->start_seq = seq;
3400                 if (after(end_seq, sp->end_seq))
3401                         sp->end_seq = end_seq;
3402                 return 1;
3403         }
3404         return 0;
3405 }
3406
3407 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3408 {
3409         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3410                 if (before(seq, tp->rcv_nxt))
3411                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3412                 else
3413                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3414
3415                 tp->rx_opt.dsack = 1;
3416                 tp->duplicate_sack[0].start_seq = seq;
3417                 tp->duplicate_sack[0].end_seq = end_seq;
3418                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3419         }
3420 }
3421
3422 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3423 {
3424         if (!tp->rx_opt.dsack)
3425                 tcp_dsack_set(tp, seq, end_seq);
3426         else
3427                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3428 }
3429
3430 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3431 {
3432         struct tcp_sock *tp = tcp_sk(sk);
3433
3434         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3435             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3436                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3437                 tcp_enter_quickack_mode(sk);
3438
3439                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3440                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3441
3442                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3443                                 end_seq = tp->rcv_nxt;
3444                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3445                 }
3446         }
3447
3448         tcp_send_ack(sk);
3449 }
3450
3451 /* These routines update the SACK block as out-of-order packets arrive or
3452  * in-order packets close up the sequence space.
3453  */
3454 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3455 {
3456         int this_sack;
3457         struct tcp_sack_block *sp = &tp->selective_acks[0];
3458         struct tcp_sack_block *swalk = sp+1;
3459
3460         /* See if the recent change to the first SACK eats into
3461          * or hits the sequence space of other SACK blocks, if so coalesce.
3462          */
3463         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3464                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3465                         int i;
3466
3467                         /* Zap SWALK, by moving every further SACK up by one slot.
3468                          * Decrease num_sacks.
3469                          */
3470                         tp->rx_opt.num_sacks--;
3471                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3472                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3473                                 sp[i] = sp[i+1];
3474                         continue;
3475                 }
3476                 this_sack++, swalk++;
3477         }
3478 }
3479
3480 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3481 {
3482         __u32 tmp;
3483
3484         tmp = sack1->start_seq;
3485         sack1->start_seq = sack2->start_seq;
3486         sack2->start_seq = tmp;
3487
3488         tmp = sack1->end_seq;
3489         sack1->end_seq = sack2->end_seq;
3490         sack2->end_seq = tmp;
3491 }
3492
3493 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3494 {
3495         struct tcp_sock *tp = tcp_sk(sk);
3496         struct tcp_sack_block *sp = &tp->selective_acks[0];
3497         int cur_sacks = tp->rx_opt.num_sacks;
3498         int this_sack;
3499
3500         if (!cur_sacks)
3501                 goto new_sack;
3502
3503         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3504                 if (tcp_sack_extend(sp, seq, end_seq)) {
3505                         /* Rotate this_sack to the first one. */
3506                         for (; this_sack>0; this_sack--, sp--)
3507                                 tcp_sack_swap(sp, sp-1);
3508                         if (cur_sacks > 1)
3509                                 tcp_sack_maybe_coalesce(tp);
3510                         return;
3511                 }
3512         }
3513
3514         /* Could not find an adjacent existing SACK, build a new one,
3515          * put it at the front, and shift everyone else down.  We
3516          * always know there is at least one SACK present already here.
3517          *
3518          * If the sack array is full, forget about the last one.
3519          */
3520         if (this_sack >= 4) {
3521                 this_sack--;
3522                 tp->rx_opt.num_sacks--;
3523                 sp--;
3524         }
3525         for (; this_sack > 0; this_sack--, sp--)
3526                 *sp = *(sp-1);
3527
3528 new_sack:
3529         /* Build the new head SACK, and we're done. */
3530         sp->start_seq = seq;
3531         sp->end_seq = end_seq;
3532         tp->rx_opt.num_sacks++;
3533         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3534 }
3535
3536 /* RCV.NXT advances, some SACKs should be eaten. */
3537
3538 static void tcp_sack_remove(struct tcp_sock *tp)
3539 {
3540         struct tcp_sack_block *sp = &tp->selective_acks[0];
3541         int num_sacks = tp->rx_opt.num_sacks;
3542         int this_sack;
3543
3544         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3545         if (skb_queue_empty(&tp->out_of_order_queue)) {
3546                 tp->rx_opt.num_sacks = 0;
3547                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3548                 return;
3549         }
3550
3551         for (this_sack = 0; this_sack < num_sacks; ) {
3552                 /* Check if the start of the sack is covered by RCV.NXT. */
3553                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3554                         int i;
3555
3556                         /* RCV.NXT must cover all the block! */
3557                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3558
3559                         /* Zap this SACK, by moving forward any other SACKS. */
3560                         for (i=this_sack+1; i < num_sacks; i++)
3561                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3562                         num_sacks--;
3563                         continue;
3564                 }
3565                 this_sack++;
3566                 sp++;
3567         }
3568         if (num_sacks != tp->rx_opt.num_sacks) {
3569                 tp->rx_opt.num_sacks = num_sacks;
3570                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3571         }
3572 }
3573
3574 /* This one checks to see if we can put data from the
3575  * out_of_order queue into the receive_queue.
3576  */
3577 static void tcp_ofo_queue(struct sock *sk)
3578 {
3579         struct tcp_sock *tp = tcp_sk(sk);
3580         __u32 dsack_high = tp->rcv_nxt;
3581         struct sk_buff *skb;
3582
3583         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3584                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3585                         break;
3586
3587                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3588                         __u32 dsack = dsack_high;
3589                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3590                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3591                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3592                 }
3593
3594                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3595                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3596                         __skb_unlink(skb, &tp->out_of_order_queue);
3597                         __kfree_skb(skb);
3598                         continue;
3599                 }
3600                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3601                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3602                            TCP_SKB_CB(skb)->end_seq);
3603
3604                 __skb_unlink(skb, &tp->out_of_order_queue);
3605                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3606                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3607                 if (tcp_hdr(skb)->fin)
3608                         tcp_fin(skb, sk, tcp_hdr(skb));
3609         }
3610 }
3611
3612 static int tcp_prune_queue(struct sock *sk);
3613
3614 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3615 {
3616         struct tcphdr *th = tcp_hdr(skb);
3617         struct tcp_sock *tp = tcp_sk(sk);
3618         int eaten = -1;
3619
3620         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3621                 goto drop;
3622
3623         __skb_pull(skb, th->doff*4);
3624
3625         TCP_ECN_accept_cwr(tp, skb);
3626
3627         if (tp->rx_opt.dsack) {
3628                 tp->rx_opt.dsack = 0;
3629                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3630                                                     4 - tp->rx_opt.tstamp_ok);
3631         }
3632
3633         /*  Queue data for delivery to the user.
3634          *  Packets in sequence go to the receive queue.
3635          *  Out of sequence packets to the out_of_order_queue.
3636          */
3637         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3638                 if (tcp_receive_window(tp) == 0)
3639                         goto out_of_window;
3640
3641                 /* Ok. In sequence. In window. */
3642                 if (tp->ucopy.task == current &&
3643                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3644                     sock_owned_by_user(sk) && !tp->urg_data) {
3645                         int chunk = min_t(unsigned int, skb->len,
3646                                                         tp->ucopy.len);
3647
3648                         __set_current_state(TASK_RUNNING);
3649
3650                         local_bh_enable();
3651                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3652                                 tp->ucopy.len -= chunk;
3653                                 tp->copied_seq += chunk;
3654                                 eaten = (chunk == skb->len && !th->fin);
3655                                 tcp_rcv_space_adjust(sk);
3656                         }
3657                         local_bh_disable();
3658                 }
3659
3660                 if (eaten <= 0) {
3661 queue_and_out:
3662                         if (eaten < 0 &&
3663                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3664                              !sk_stream_rmem_schedule(sk, skb))) {
3665                                 if (tcp_prune_queue(sk) < 0 ||
3666                                     !sk_stream_rmem_schedule(sk, skb))
3667                                         goto drop;
3668                         }
3669                         sk_stream_set_owner_r(skb, sk);
3670                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3671                 }
3672                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3673                 if (skb->len)
3674                         tcp_event_data_recv(sk, skb);
3675                 if (th->fin)
3676                         tcp_fin(skb, sk, th);
3677
3678                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3679                         tcp_ofo_queue(sk);
3680
3681                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3682                          * gap in queue is filled.
3683                          */
3684                         if (skb_queue_empty(&tp->out_of_order_queue))
3685                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3686                 }
3687
3688                 if (tp->rx_opt.num_sacks)
3689                         tcp_sack_remove(tp);
3690
3691                 tcp_fast_path_check(sk);
3692
3693                 if (eaten > 0)
3694                         __kfree_skb(skb);
3695                 else if (!sock_flag(sk, SOCK_DEAD))
3696                         sk->sk_data_ready(sk, 0);
3697                 return;
3698         }
3699
3700         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3701                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3702                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3703                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3704
3705 out_of_window:
3706                 tcp_enter_quickack_mode(sk);
3707                 inet_csk_schedule_ack(sk);
3708 drop:
3709                 __kfree_skb(skb);
3710                 return;
3711         }
3712
3713         /* Out of window. F.e. zero window probe. */
3714         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3715                 goto out_of_window;
3716
3717         tcp_enter_quickack_mode(sk);
3718
3719         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3720                 /* Partial packet, seq < rcv_next < end_seq */
3721                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3722                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3723                            TCP_SKB_CB(skb)->end_seq);
3724
3725                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3726
3727                 /* If window is closed, drop tail of packet. But after
3728                  * remembering D-SACK for its head made in previous line.
3729                  */
3730                 if (!tcp_receive_window(tp))
3731                         goto out_of_window;
3732                 goto queue_and_out;
3733         }
3734
3735         TCP_ECN_check_ce(tp, skb);
3736
3737         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3738             !sk_stream_rmem_schedule(sk, skb)) {
3739                 if (tcp_prune_queue(sk) < 0 ||
3740                     !sk_stream_rmem_schedule(sk, skb))
3741                         goto drop;
3742         }
3743
3744         /* Disable header prediction. */
3745         tp->pred_flags = 0;
3746         inet_csk_schedule_ack(sk);
3747
3748         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3749                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3750
3751         sk_stream_set_owner_r(skb, sk);
3752
3753         if (!skb_peek(&tp->out_of_order_queue)) {
3754                 /* Initial out of order segment, build 1 SACK. */
3755                 if (tcp_is_sack(tp)) {
3756                         tp->rx_opt.num_sacks = 1;
3757                         tp->rx_opt.dsack     = 0;
3758                         tp->rx_opt.eff_sacks = 1;
3759                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3760                         tp->selective_acks[0].end_seq =
3761                                                 TCP_SKB_CB(skb)->end_seq;
3762                 }
3763                 __skb_queue_head(&tp->out_of_order_queue,skb);
3764         } else {
3765                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3766                 u32 seq = TCP_SKB_CB(skb)->seq;
3767                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3768
3769                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3770                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3771
3772                         if (!tp->rx_opt.num_sacks ||
3773                             tp->selective_acks[0].end_seq != seq)
3774                                 goto add_sack;
3775
3776                         /* Common case: data arrive in order after hole. */
3777                         tp->selective_acks[0].end_seq = end_seq;
3778                         return;
3779                 }
3780
3781                 /* Find place to insert this segment. */
3782                 do {
3783                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3784                                 break;
3785                 } while ((skb1 = skb1->prev) !=
3786                          (struct sk_buff*)&tp->out_of_order_queue);
3787
3788                 /* Do skb overlap to previous one? */
3789                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3790                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3791                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3792                                 /* All the bits are present. Drop. */
3793                                 __kfree_skb(skb);
3794                                 tcp_dsack_set(tp, seq, end_seq);
3795                                 goto add_sack;
3796                         }
3797                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3798                                 /* Partial overlap. */
3799                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3800                         } else {
3801                                 skb1 = skb1->prev;
3802                         }
3803                 }
3804                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3805
3806                 /* And clean segments covered by new one as whole. */
3807                 while ((skb1 = skb->next) !=
3808                        (struct sk_buff*)&tp->out_of_order_queue &&
3809                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3810                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3811                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3812                                break;
3813                        }
3814                        __skb_unlink(skb1, &tp->out_of_order_queue);
3815                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3816                        __kfree_skb(skb1);
3817                 }
3818
3819 add_sack:
3820                 if (tcp_is_sack(tp))
3821                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3822         }
3823 }
3824
3825 /* Collapse contiguous sequence of skbs head..tail with
3826  * sequence numbers start..end.
3827  * Segments with FIN/SYN are not collapsed (only because this
3828  * simplifies code)
3829  */
3830 static void
3831 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3832              struct sk_buff *head, struct sk_buff *tail,
3833              u32 start, u32 end)
3834 {
3835         struct sk_buff *skb;
3836
3837         /* First, check that queue is collapsible and find
3838          * the point where collapsing can be useful. */
3839         for (skb = head; skb != tail; ) {
3840                 /* No new bits? It is possible on ofo queue. */
3841                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3842                         struct sk_buff *next = skb->next;
3843                         __skb_unlink(skb, list);
3844                         __kfree_skb(skb);
3845                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3846                         skb = next;
3847                         continue;
3848                 }
3849
3850                 /* The first skb to collapse is:
3851                  * - not SYN/FIN and
3852                  * - bloated or contains data before "start" or
3853                  *   overlaps to the next one.
3854                  */
3855                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3856                     (tcp_win_from_space(skb->truesize) > skb->len ||
3857                      before(TCP_SKB_CB(skb)->seq, start) ||
3858                      (skb->next != tail &&
3859                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3860                         break;
3861
3862                 /* Decided to skip this, advance start seq. */
3863                 start = TCP_SKB_CB(skb)->end_seq;
3864                 skb = skb->next;
3865         }
3866         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3867                 return;
3868
3869         while (before(start, end)) {
3870                 struct sk_buff *nskb;
3871                 int header = skb_headroom(skb);
3872                 int copy = SKB_MAX_ORDER(header, 0);
3873
3874                 /* Too big header? This can happen with IPv6. */
3875                 if (copy < 0)
3876                         return;
3877                 if (end-start < copy)
3878                         copy = end-start;
3879                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3880                 if (!nskb)
3881                         return;
3882
3883                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3884                 skb_set_network_header(nskb, (skb_network_header(skb) -
3885                                               skb->head));
3886                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3887                                                 skb->head));
3888                 skb_reserve(nskb, header);
3889                 memcpy(nskb->head, skb->head, header);
3890                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3891                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3892                 __skb_insert(nskb, skb->prev, skb, list);
3893                 sk_stream_set_owner_r(nskb, sk);
3894
3895                 /* Copy data, releasing collapsed skbs. */
3896                 while (copy > 0) {
3897                         int offset = start - TCP_SKB_CB(skb)->seq;
3898                         int size = TCP_SKB_CB(skb)->end_seq - start;
3899
3900                         BUG_ON(offset < 0);
3901                         if (size > 0) {
3902                                 size = min(copy, size);
3903                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3904                                         BUG();
3905                                 TCP_SKB_CB(nskb)->end_seq += size;
3906                                 copy -= size;
3907                                 start += size;
3908                         }
3909                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3910                                 struct sk_buff *next = skb->next;
3911                                 __skb_unlink(skb, list);
3912                                 __kfree_skb(skb);
3913                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3914                                 skb = next;
3915                                 if (skb == tail ||
3916                                     tcp_hdr(skb)->syn ||
3917                                     tcp_hdr(skb)->fin)
3918                                         return;
3919                         }
3920                 }
3921         }
3922 }
3923
3924 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3925  * and tcp_collapse() them until all the queue is collapsed.
3926  */
3927 static void tcp_collapse_ofo_queue(struct sock *sk)
3928 {
3929         struct tcp_sock *tp = tcp_sk(sk);
3930         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3931         struct sk_buff *head;
3932         u32 start, end;
3933
3934         if (skb == NULL)
3935                 return;
3936
3937         start = TCP_SKB_CB(skb)->seq;
3938         end = TCP_SKB_CB(skb)->end_seq;
3939         head = skb;
3940
3941         for (;;) {
3942                 skb = skb->next;
3943
3944                 /* Segment is terminated when we see gap or when
3945                  * we are at the end of all the queue. */
3946                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3947                     after(TCP_SKB_CB(skb)->seq, end) ||
3948                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3949                         tcp_collapse(sk, &tp->out_of_order_queue,
3950                                      head, skb, start, end);
3951                         head = skb;
3952                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3953                                 break;
3954                         /* Start new segment */
3955                         start = TCP_SKB_CB(skb)->seq;
3956                         end = TCP_SKB_CB(skb)->end_seq;
3957                 } else {
3958                         if (before(TCP_SKB_CB(skb)->seq, start))
3959                                 start = TCP_SKB_CB(skb)->seq;
3960                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3961                                 end = TCP_SKB_CB(skb)->end_seq;
3962                 }
3963         }
3964 }
3965
3966 /* Reduce allocated memory if we can, trying to get
3967  * the socket within its memory limits again.
3968  *
3969  * Return less than zero if we should start dropping frames
3970  * until the socket owning process reads some of the data
3971  * to stabilize the situation.
3972  */
3973 static int tcp_prune_queue(struct sock *sk)
3974 {
3975         struct tcp_sock *tp = tcp_sk(sk);
3976
3977         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3978
3979         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3980
3981         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3982                 tcp_clamp_window(sk);
3983         else if (tcp_memory_pressure)
3984                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3985
3986         tcp_collapse_ofo_queue(sk);
3987         tcp_collapse(sk, &sk->sk_receive_queue,
3988                      sk->sk_receive_queue.next,
3989                      (struct sk_buff*)&sk->sk_receive_queue,
3990                      tp->copied_seq, tp->rcv_nxt);
3991         sk_stream_mem_reclaim(sk);
3992
3993         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3994                 return 0;
3995
3996         /* Collapsing did not help, destructive actions follow.
3997          * This must not ever occur. */
3998
3999         /* First, purge the out_of_order queue. */
4000         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4001                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
4002                 __skb_queue_purge(&tp->out_of_order_queue);
4003
4004                 /* Reset SACK state.  A conforming SACK implementation will
4005                  * do the same at a timeout based retransmit.  When a connection
4006                  * is in a sad state like this, we care only about integrity
4007                  * of the connection not performance.
4008                  */
4009                 if (tcp_is_sack(tp))
4010                         tcp_sack_reset(&tp->rx_opt);
4011                 sk_stream_mem_reclaim(sk);
4012         }
4013
4014         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4015                 return 0;
4016
4017         /* If we are really being abused, tell the caller to silently
4018          * drop receive data on the floor.  It will get retransmitted
4019          * and hopefully then we'll have sufficient space.
4020          */
4021         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4022
4023         /* Massive buffer overcommit. */
4024         tp->pred_flags = 0;
4025         return -1;
4026 }
4027
4028
4029 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4030  * As additional protections, we do not touch cwnd in retransmission phases,
4031  * and if application hit its sndbuf limit recently.
4032  */
4033 void tcp_cwnd_application_limited(struct sock *sk)
4034 {
4035         struct tcp_sock *tp = tcp_sk(sk);
4036
4037         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4038             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4039                 /* Limited by application or receiver window. */
4040                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4041                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4042                 if (win_used < tp->snd_cwnd) {
4043                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4044                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4045                 }
4046                 tp->snd_cwnd_used = 0;
4047         }
4048         tp->snd_cwnd_stamp = tcp_time_stamp;
4049 }
4050
4051 static int tcp_should_expand_sndbuf(struct sock *sk)
4052 {
4053         struct tcp_sock *tp = tcp_sk(sk);
4054
4055         /* If the user specified a specific send buffer setting, do
4056          * not modify it.
4057          */
4058         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4059                 return 0;
4060
4061         /* If we are under global TCP memory pressure, do not expand.  */
4062         if (tcp_memory_pressure)
4063                 return 0;
4064
4065         /* If we are under soft global TCP memory pressure, do not expand.  */
4066         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4067                 return 0;
4068
4069         /* If we filled the congestion window, do not expand.  */
4070         if (tp->packets_out >= tp->snd_cwnd)
4071                 return 0;
4072
4073         return 1;
4074 }
4075
4076 /* When incoming ACK allowed to free some skb from write_queue,
4077  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4078  * on the exit from tcp input handler.
4079  *
4080  * PROBLEM: sndbuf expansion does not work well with largesend.
4081  */
4082 static void tcp_new_space(struct sock *sk)
4083 {
4084         struct tcp_sock *tp = tcp_sk(sk);
4085
4086         if (tcp_should_expand_sndbuf(sk)) {
4087                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4088                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4089                     demanded = max_t(unsigned int, tp->snd_cwnd,
4090                                                    tp->reordering + 1);
4091                 sndmem *= 2*demanded;
4092                 if (sndmem > sk->sk_sndbuf)
4093                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4094                 tp->snd_cwnd_stamp = tcp_time_stamp;
4095         }
4096
4097         sk->sk_write_space(sk);
4098 }
4099
4100 static void tcp_check_space(struct sock *sk)
4101 {
4102         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4103                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4104                 if (sk->sk_socket &&
4105                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4106                         tcp_new_space(sk);
4107         }
4108 }
4109
4110 static inline void tcp_data_snd_check(struct sock *sk)
4111 {
4112         tcp_push_pending_frames(sk);
4113         tcp_check_space(sk);
4114 }
4115
4116 /*
4117  * Check if sending an ack is needed.
4118  */
4119 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4120 {
4121         struct tcp_sock *tp = tcp_sk(sk);
4122
4123             /* More than one full frame received... */
4124         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4125              /* ... and right edge of window advances far enough.
4126               * (tcp_recvmsg() will send ACK otherwise). Or...
4127               */
4128              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4129             /* We ACK each frame or... */
4130             tcp_in_quickack_mode(sk) ||
4131             /* We have out of order data. */
4132             (ofo_possible &&
4133              skb_peek(&tp->out_of_order_queue))) {
4134                 /* Then ack it now */
4135                 tcp_send_ack(sk);
4136         } else {
4137                 /* Else, send delayed ack. */
4138                 tcp_send_delayed_ack(sk);
4139         }
4140 }
4141
4142 static inline void tcp_ack_snd_check(struct sock *sk)
4143 {
4144         if (!inet_csk_ack_scheduled(sk)) {
4145                 /* We sent a data segment already. */
4146                 return;
4147         }
4148         __tcp_ack_snd_check(sk, 1);
4149 }
4150
4151 /*
4152  *      This routine is only called when we have urgent data
4153  *      signaled. Its the 'slow' part of tcp_urg. It could be
4154  *      moved inline now as tcp_urg is only called from one
4155  *      place. We handle URGent data wrong. We have to - as
4156  *      BSD still doesn't use the correction from RFC961.
4157  *      For 1003.1g we should support a new option TCP_STDURG to permit
4158  *      either form (or just set the sysctl tcp_stdurg).
4159  */
4160
4161 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4162 {
4163         struct tcp_sock *tp = tcp_sk(sk);
4164         u32 ptr = ntohs(th->urg_ptr);
4165
4166         if (ptr && !sysctl_tcp_stdurg)
4167                 ptr--;
4168         ptr += ntohl(th->seq);
4169
4170         /* Ignore urgent data that we've already seen and read. */
4171         if (after(tp->copied_seq, ptr))
4172                 return;
4173
4174         /* Do not replay urg ptr.
4175          *
4176          * NOTE: interesting situation not covered by specs.
4177          * Misbehaving sender may send urg ptr, pointing to segment,
4178          * which we already have in ofo queue. We are not able to fetch
4179          * such data and will stay in TCP_URG_NOTYET until will be eaten
4180          * by recvmsg(). Seems, we are not obliged to handle such wicked
4181          * situations. But it is worth to think about possibility of some
4182          * DoSes using some hypothetical application level deadlock.
4183          */
4184         if (before(ptr, tp->rcv_nxt))
4185                 return;
4186
4187         /* Do we already have a newer (or duplicate) urgent pointer? */
4188         if (tp->urg_data && !after(ptr, tp->urg_seq))
4189                 return;
4190
4191         /* Tell the world about our new urgent pointer. */
4192         sk_send_sigurg(sk);
4193
4194         /* We may be adding urgent data when the last byte read was
4195          * urgent. To do this requires some care. We cannot just ignore
4196          * tp->copied_seq since we would read the last urgent byte again
4197          * as data, nor can we alter copied_seq until this data arrives
4198          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4199          *
4200          * NOTE. Double Dutch. Rendering to plain English: author of comment
4201          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4202          * and expect that both A and B disappear from stream. This is _wrong_.
4203          * Though this happens in BSD with high probability, this is occasional.
4204          * Any application relying on this is buggy. Note also, that fix "works"
4205          * only in this artificial test. Insert some normal data between A and B and we will
4206          * decline of BSD again. Verdict: it is better to remove to trap
4207          * buggy users.
4208          */
4209         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4210             !sock_flag(sk, SOCK_URGINLINE) &&
4211             tp->copied_seq != tp->rcv_nxt) {
4212                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4213                 tp->copied_seq++;
4214                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4215                         __skb_unlink(skb, &sk->sk_receive_queue);
4216                         __kfree_skb(skb);
4217                 }
4218         }
4219
4220         tp->urg_data   = TCP_URG_NOTYET;
4221         tp->urg_seq    = ptr;
4222
4223         /* Disable header prediction. */
4224         tp->pred_flags = 0;
4225 }
4226
4227 /* This is the 'fast' part of urgent handling. */
4228 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4229 {
4230         struct tcp_sock *tp = tcp_sk(sk);
4231
4232         /* Check if we get a new urgent pointer - normally not. */
4233         if (th->urg)
4234                 tcp_check_urg(sk,th);
4235
4236         /* Do we wait for any urgent data? - normally not... */
4237         if (tp->urg_data == TCP_URG_NOTYET) {
4238                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4239                           th->syn;
4240
4241                 /* Is the urgent pointer pointing into this packet? */
4242                 if (ptr < skb->len) {
4243                         u8 tmp;
4244                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4245                                 BUG();
4246                         tp->urg_data = TCP_URG_VALID | tmp;
4247                         if (!sock_flag(sk, SOCK_DEAD))
4248                                 sk->sk_data_ready(sk, 0);
4249                 }
4250         }
4251 }
4252
4253 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4254 {
4255         struct tcp_sock *tp = tcp_sk(sk);
4256         int chunk = skb->len - hlen;
4257         int err;
4258
4259         local_bh_enable();
4260         if (skb_csum_unnecessary(skb))
4261                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4262         else
4263                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4264                                                        tp->ucopy.iov);
4265
4266         if (!err) {
4267                 tp->ucopy.len -= chunk;
4268                 tp->copied_seq += chunk;
4269                 tcp_rcv_space_adjust(sk);
4270         }
4271
4272         local_bh_disable();
4273         return err;
4274 }
4275
4276 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4277 {
4278         __sum16 result;
4279
4280         if (sock_owned_by_user(sk)) {
4281                 local_bh_enable();
4282                 result = __tcp_checksum_complete(skb);
4283                 local_bh_disable();
4284         } else {
4285                 result = __tcp_checksum_complete(skb);
4286         }
4287         return result;
4288 }
4289
4290 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4291 {
4292         return !skb_csum_unnecessary(skb) &&
4293                 __tcp_checksum_complete_user(sk, skb);
4294 }
4295
4296 #ifdef CONFIG_NET_DMA
4297 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4298 {
4299         struct tcp_sock *tp = tcp_sk(sk);
4300         int chunk = skb->len - hlen;
4301         int dma_cookie;
4302         int copied_early = 0;
4303
4304         if (tp->ucopy.wakeup)
4305                 return 0;
4306
4307         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4308                 tp->ucopy.dma_chan = get_softnet_dma();
4309
4310         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4311
4312                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4313                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4314
4315                 if (dma_cookie < 0)
4316                         goto out;
4317
4318                 tp->ucopy.dma_cookie = dma_cookie;
4319                 copied_early = 1;
4320
4321                 tp->ucopy.len -= chunk;
4322                 tp->copied_seq += chunk;
4323                 tcp_rcv_space_adjust(sk);
4324
4325                 if ((tp->ucopy.len == 0) ||
4326                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4327                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4328                         tp->ucopy.wakeup = 1;
4329                         sk->sk_data_ready(sk, 0);
4330                 }
4331         } else if (chunk > 0) {
4332                 tp->ucopy.wakeup = 1;
4333                 sk->sk_data_ready(sk, 0);
4334         }
4335 out:
4336         return copied_early;
4337 }
4338 #endif /* CONFIG_NET_DMA */
4339
4340 /*
4341  *      TCP receive function for the ESTABLISHED state.
4342  *
4343  *      It is split into a fast path and a slow path. The fast path is
4344  *      disabled when:
4345  *      - A zero window was announced from us - zero window probing
4346  *        is only handled properly in the slow path.
4347  *      - Out of order segments arrived.
4348  *      - Urgent data is expected.
4349  *      - There is no buffer space left
4350  *      - Unexpected TCP flags/window values/header lengths are received
4351  *        (detected by checking the TCP header against pred_flags)
4352  *      - Data is sent in both directions. Fast path only supports pure senders
4353  *        or pure receivers (this means either the sequence number or the ack
4354  *        value must stay constant)
4355  *      - Unexpected TCP option.
4356  *
4357  *      When these conditions are not satisfied it drops into a standard
4358  *      receive procedure patterned after RFC793 to handle all cases.
4359  *      The first three cases are guaranteed by proper pred_flags setting,
4360  *      the rest is checked inline. Fast processing is turned on in
4361  *      tcp_data_queue when everything is OK.
4362  */
4363 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4364                         struct tcphdr *th, unsigned len)
4365 {
4366         struct tcp_sock *tp = tcp_sk(sk);
4367
4368         /*
4369          *      Header prediction.
4370          *      The code loosely follows the one in the famous
4371          *      "30 instruction TCP receive" Van Jacobson mail.
4372          *
4373          *      Van's trick is to deposit buffers into socket queue
4374          *      on a device interrupt, to call tcp_recv function
4375          *      on the receive process context and checksum and copy
4376          *      the buffer to user space. smart...
4377          *
4378          *      Our current scheme is not silly either but we take the
4379          *      extra cost of the net_bh soft interrupt processing...
4380          *      We do checksum and copy also but from device to kernel.
4381          */
4382
4383         tp->rx_opt.saw_tstamp = 0;
4384
4385         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4386          *      if header_prediction is to be made
4387          *      'S' will always be tp->tcp_header_len >> 2
4388          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4389          *  turn it off (when there are holes in the receive
4390          *       space for instance)
4391          *      PSH flag is ignored.
4392          */
4393
4394         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4395                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4396                 int tcp_header_len = tp->tcp_header_len;
4397
4398                 /* Timestamp header prediction: tcp_header_len
4399                  * is automatically equal to th->doff*4 due to pred_flags
4400                  * match.
4401                  */
4402
4403                 /* Check timestamp */
4404                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4405                         __be32 *ptr = (__be32 *)(th + 1);
4406
4407                         /* No? Slow path! */
4408                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4409                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4410                                 goto slow_path;
4411
4412                         tp->rx_opt.saw_tstamp = 1;
4413                         ++ptr;
4414                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4415                         ++ptr;
4416                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4417
4418                         /* If PAWS failed, check it more carefully in slow path */
4419                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4420                                 goto slow_path;
4421
4422                         /* DO NOT update ts_recent here, if checksum fails
4423                          * and timestamp was corrupted part, it will result
4424                          * in a hung connection since we will drop all
4425                          * future packets due to the PAWS test.
4426                          */
4427                 }
4428
4429                 if (len <= tcp_header_len) {
4430                         /* Bulk data transfer: sender */
4431                         if (len == tcp_header_len) {
4432                                 /* Predicted packet is in window by definition.
4433                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4434                                  * Hence, check seq<=rcv_wup reduces to:
4435                                  */
4436                                 if (tcp_header_len ==
4437                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4438                                     tp->rcv_nxt == tp->rcv_wup)
4439                                         tcp_store_ts_recent(tp);
4440
4441                                 /* We know that such packets are checksummed
4442                                  * on entry.
4443                                  */
4444                                 tcp_ack(sk, skb, 0);
4445                                 __kfree_skb(skb);
4446                                 tcp_data_snd_check(sk);
4447                                 return 0;
4448                         } else { /* Header too small */
4449                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4450                                 goto discard;
4451                         }
4452                 } else {
4453                         int eaten = 0;
4454                         int copied_early = 0;
4455
4456                         if (tp->copied_seq == tp->rcv_nxt &&
4457                             len - tcp_header_len <= tp->ucopy.len) {
4458 #ifdef CONFIG_NET_DMA
4459                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4460                                         copied_early = 1;
4461                                         eaten = 1;
4462                                 }
4463 #endif
4464                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4465                                         __set_current_state(TASK_RUNNING);
4466
4467                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4468                                                 eaten = 1;
4469                                 }
4470                                 if (eaten) {
4471                                         /* Predicted packet is in window by definition.
4472                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4473                                          * Hence, check seq<=rcv_wup reduces to:
4474                                          */
4475                                         if (tcp_header_len ==
4476                                             (sizeof(struct tcphdr) +
4477                                              TCPOLEN_TSTAMP_ALIGNED) &&
4478                                             tp->rcv_nxt == tp->rcv_wup)
4479                                                 tcp_store_ts_recent(tp);
4480
4481                                         tcp_rcv_rtt_measure_ts(sk, skb);
4482
4483                                         __skb_pull(skb, tcp_header_len);
4484                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4485                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4486                                 }
4487                                 if (copied_early)
4488                                         tcp_cleanup_rbuf(sk, skb->len);
4489                         }
4490                         if (!eaten) {
4491                                 if (tcp_checksum_complete_user(sk, skb))
4492                                         goto csum_error;
4493
4494                                 /* Predicted packet is in window by definition.
4495                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4496                                  * Hence, check seq<=rcv_wup reduces to:
4497                                  */
4498                                 if (tcp_header_len ==
4499                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4500                                     tp->rcv_nxt == tp->rcv_wup)
4501                                         tcp_store_ts_recent(tp);
4502
4503                                 tcp_rcv_rtt_measure_ts(sk, skb);
4504
4505                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4506                                         goto step5;
4507
4508                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4509
4510                                 /* Bulk data transfer: receiver */
4511                                 __skb_pull(skb,tcp_header_len);
4512                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4513                                 sk_stream_set_owner_r(skb, sk);
4514                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4515                         }
4516
4517                         tcp_event_data_recv(sk, skb);
4518
4519                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4520                                 /* Well, only one small jumplet in fast path... */
4521                                 tcp_ack(sk, skb, FLAG_DATA);
4522                                 tcp_data_snd_check(sk);
4523                                 if (!inet_csk_ack_scheduled(sk))
4524                                         goto no_ack;
4525                         }
4526
4527                         __tcp_ack_snd_check(sk, 0);
4528 no_ack:
4529 #ifdef CONFIG_NET_DMA
4530                         if (copied_early)
4531                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4532                         else
4533 #endif
4534                         if (eaten)
4535                                 __kfree_skb(skb);
4536                         else
4537                                 sk->sk_data_ready(sk, 0);
4538                         return 0;
4539                 }
4540         }
4541
4542 slow_path:
4543         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4544                 goto csum_error;
4545
4546         /*
4547          * RFC1323: H1. Apply PAWS check first.
4548          */
4549         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4550             tcp_paws_discard(sk, skb)) {
4551                 if (!th->rst) {
4552                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4553                         tcp_send_dupack(sk, skb);
4554                         goto discard;
4555                 }
4556                 /* Resets are accepted even if PAWS failed.
4557
4558                    ts_recent update must be made after we are sure
4559                    that the packet is in window.
4560                  */
4561         }
4562
4563         /*
4564          *      Standard slow path.
4565          */
4566
4567         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4568                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4569                  * (RST) segments are validated by checking their SEQ-fields."
4570                  * And page 69: "If an incoming segment is not acceptable,
4571                  * an acknowledgment should be sent in reply (unless the RST bit
4572                  * is set, if so drop the segment and return)".
4573                  */
4574                 if (!th->rst)
4575                         tcp_send_dupack(sk, skb);
4576                 goto discard;
4577         }
4578
4579         if (th->rst) {
4580                 tcp_reset(sk);
4581                 goto discard;
4582         }
4583
4584         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4585
4586         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4587                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4588                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4589                 tcp_reset(sk);
4590                 return 1;
4591         }
4592
4593 step5:
4594         if (th->ack)
4595                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4596
4597         tcp_rcv_rtt_measure_ts(sk, skb);
4598
4599         /* Process urgent data. */
4600         tcp_urg(sk, skb, th);
4601
4602         /* step 7: process the segment text */
4603         tcp_data_queue(sk, skb);
4604
4605         tcp_data_snd_check(sk);
4606         tcp_ack_snd_check(sk);
4607         return 0;
4608
4609 csum_error:
4610         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4611
4612 discard:
4613         __kfree_skb(skb);
4614         return 0;
4615 }
4616
4617 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4618                                          struct tcphdr *th, unsigned len)
4619 {
4620         struct tcp_sock *tp = tcp_sk(sk);
4621         struct inet_connection_sock *icsk = inet_csk(sk);
4622         int saved_clamp = tp->rx_opt.mss_clamp;
4623
4624         tcp_parse_options(skb, &tp->rx_opt, 0);
4625
4626         if (th->ack) {
4627                 /* rfc793:
4628                  * "If the state is SYN-SENT then
4629                  *    first check the ACK bit
4630                  *      If the ACK bit is set
4631                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4632                  *        a reset (unless the RST bit is set, if so drop
4633                  *        the segment and return)"
4634                  *
4635                  *  We do not send data with SYN, so that RFC-correct
4636                  *  test reduces to:
4637                  */
4638                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4639                         goto reset_and_undo;
4640
4641                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4642                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4643                              tcp_time_stamp)) {
4644                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4645                         goto reset_and_undo;
4646                 }
4647
4648                 /* Now ACK is acceptable.
4649                  *
4650                  * "If the RST bit is set
4651                  *    If the ACK was acceptable then signal the user "error:
4652                  *    connection reset", drop the segment, enter CLOSED state,
4653                  *    delete TCB, and return."
4654                  */
4655
4656                 if (th->rst) {
4657                         tcp_reset(sk);
4658                         goto discard;
4659                 }
4660
4661                 /* rfc793:
4662                  *   "fifth, if neither of the SYN or RST bits is set then
4663                  *    drop the segment and return."
4664                  *
4665                  *    See note below!
4666                  *                                        --ANK(990513)
4667                  */
4668                 if (!th->syn)
4669                         goto discard_and_undo;
4670
4671                 /* rfc793:
4672                  *   "If the SYN bit is on ...
4673                  *    are acceptable then ...
4674                  *    (our SYN has been ACKed), change the connection
4675                  *    state to ESTABLISHED..."
4676                  */
4677
4678                 TCP_ECN_rcv_synack(tp, th);
4679
4680                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4681                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4682
4683                 /* Ok.. it's good. Set up sequence numbers and
4684                  * move to established.
4685                  */
4686                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4687                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4688
4689                 /* RFC1323: The window in SYN & SYN/ACK segments is
4690                  * never scaled.
4691                  */
4692                 tp->snd_wnd = ntohs(th->window);
4693                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4694
4695                 if (!tp->rx_opt.wscale_ok) {
4696                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4697                         tp->window_clamp = min(tp->window_clamp, 65535U);
4698                 }
4699
4700                 if (tp->rx_opt.saw_tstamp) {
4701                         tp->rx_opt.tstamp_ok       = 1;
4702                         tp->tcp_header_len =
4703                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4704                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4705                         tcp_store_ts_recent(tp);
4706                 } else {
4707                         tp->tcp_header_len = sizeof(struct tcphdr);
4708                 }
4709
4710                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4711                         tcp_enable_fack(tp);
4712
4713                 tcp_mtup_init(sk);
4714                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4715                 tcp_initialize_rcv_mss(sk);
4716
4717                 /* Remember, tcp_poll() does not lock socket!
4718                  * Change state from SYN-SENT only after copied_seq
4719                  * is initialized. */
4720                 tp->copied_seq = tp->rcv_nxt;
4721                 smp_mb();
4722                 tcp_set_state(sk, TCP_ESTABLISHED);
4723
4724                 security_inet_conn_established(sk, skb);
4725
4726                 /* Make sure socket is routed, for correct metrics.  */
4727                 icsk->icsk_af_ops->rebuild_header(sk);
4728
4729                 tcp_init_metrics(sk);
4730
4731                 tcp_init_congestion_control(sk);
4732
4733                 /* Prevent spurious tcp_cwnd_restart() on first data
4734                  * packet.
4735                  */
4736                 tp->lsndtime = tcp_time_stamp;
4737
4738                 tcp_init_buffer_space(sk);
4739
4740                 if (sock_flag(sk, SOCK_KEEPOPEN))
4741                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4742
4743                 if (!tp->rx_opt.snd_wscale)
4744                         __tcp_fast_path_on(tp, tp->snd_wnd);
4745                 else
4746                         tp->pred_flags = 0;
4747
4748                 if (!sock_flag(sk, SOCK_DEAD)) {
4749                         sk->sk_state_change(sk);
4750                         sk_wake_async(sk, 0, POLL_OUT);
4751                 }
4752
4753                 if (sk->sk_write_pending ||
4754                     icsk->icsk_accept_queue.rskq_defer_accept ||
4755                     icsk->icsk_ack.pingpong) {
4756                         /* Save one ACK. Data will be ready after
4757                          * several ticks, if write_pending is set.
4758                          *
4759                          * It may be deleted, but with this feature tcpdumps
4760                          * look so _wonderfully_ clever, that I was not able
4761                          * to stand against the temptation 8)     --ANK
4762                          */
4763                         inet_csk_schedule_ack(sk);
4764                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4765                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4766                         tcp_incr_quickack(sk);
4767                         tcp_enter_quickack_mode(sk);
4768                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4769                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4770
4771 discard:
4772                         __kfree_skb(skb);
4773                         return 0;
4774                 } else {
4775                         tcp_send_ack(sk);
4776                 }
4777                 return -1;
4778         }
4779
4780         /* No ACK in the segment */
4781
4782         if (th->rst) {
4783                 /* rfc793:
4784                  * "If the RST bit is set
4785                  *
4786                  *      Otherwise (no ACK) drop the segment and return."
4787                  */
4788
4789                 goto discard_and_undo;
4790         }
4791
4792         /* PAWS check. */
4793         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4794                 goto discard_and_undo;
4795
4796         if (th->syn) {
4797                 /* We see SYN without ACK. It is attempt of
4798                  * simultaneous connect with crossed SYNs.
4799                  * Particularly, it can be connect to self.
4800                  */
4801                 tcp_set_state(sk, TCP_SYN_RECV);
4802
4803                 if (tp->rx_opt.saw_tstamp) {
4804                         tp->rx_opt.tstamp_ok = 1;
4805                         tcp_store_ts_recent(tp);
4806                         tp->tcp_header_len =
4807                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4808                 } else {
4809                         tp->tcp_header_len = sizeof(struct tcphdr);
4810                 }
4811
4812                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4813                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4814
4815                 /* RFC1323: The window in SYN & SYN/ACK segments is
4816                  * never scaled.
4817                  */
4818                 tp->snd_wnd    = ntohs(th->window);
4819                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4820                 tp->max_window = tp->snd_wnd;
4821
4822                 TCP_ECN_rcv_syn(tp, th);
4823
4824                 tcp_mtup_init(sk);
4825                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4826                 tcp_initialize_rcv_mss(sk);
4827
4828
4829                 tcp_send_synack(sk);
4830 #if 0
4831                 /* Note, we could accept data and URG from this segment.
4832                  * There are no obstacles to make this.
4833                  *
4834                  * However, if we ignore data in ACKless segments sometimes,
4835                  * we have no reasons to accept it sometimes.
4836                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4837                  * is not flawless. So, discard packet for sanity.
4838                  * Uncomment this return to process the data.
4839                  */
4840                 return -1;
4841 #else
4842                 goto discard;
4843 #endif
4844         }
4845         /* "fifth, if neither of the SYN or RST bits is set then
4846          * drop the segment and return."
4847          */
4848
4849 discard_and_undo:
4850         tcp_clear_options(&tp->rx_opt);
4851         tp->rx_opt.mss_clamp = saved_clamp;
4852         goto discard;
4853
4854 reset_and_undo:
4855         tcp_clear_options(&tp->rx_opt);
4856         tp->rx_opt.mss_clamp = saved_clamp;
4857         return 1;
4858 }
4859
4860
4861 /*
4862  *      This function implements the receiving procedure of RFC 793 for
4863  *      all states except ESTABLISHED and TIME_WAIT.
4864  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4865  *      address independent.
4866  */
4867
4868 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4869                           struct tcphdr *th, unsigned len)
4870 {
4871         struct tcp_sock *tp = tcp_sk(sk);
4872         struct inet_connection_sock *icsk = inet_csk(sk);
4873         int queued = 0;
4874
4875         tp->rx_opt.saw_tstamp = 0;
4876
4877         switch (sk->sk_state) {
4878         case TCP_CLOSE:
4879                 goto discard;
4880
4881         case TCP_LISTEN:
4882                 if (th->ack)
4883                         return 1;
4884
4885                 if (th->rst)
4886                         goto discard;
4887
4888                 if (th->syn) {
4889                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4890                                 return 1;
4891
4892                         /* Now we have several options: In theory there is
4893                          * nothing else in the frame. KA9Q has an option to
4894                          * send data with the syn, BSD accepts data with the
4895                          * syn up to the [to be] advertised window and
4896                          * Solaris 2.1 gives you a protocol error. For now
4897                          * we just ignore it, that fits the spec precisely
4898                          * and avoids incompatibilities. It would be nice in
4899                          * future to drop through and process the data.
4900                          *
4901                          * Now that TTCP is starting to be used we ought to
4902                          * queue this data.
4903                          * But, this leaves one open to an easy denial of
4904                          * service attack, and SYN cookies can't defend
4905                          * against this problem. So, we drop the data
4906                          * in the interest of security over speed unless
4907                          * it's still in use.
4908                          */
4909                         kfree_skb(skb);
4910                         return 0;
4911                 }
4912                 goto discard;
4913
4914         case TCP_SYN_SENT:
4915                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4916                 if (queued >= 0)
4917                         return queued;
4918
4919                 /* Do step6 onward by hand. */
4920                 tcp_urg(sk, skb, th);
4921                 __kfree_skb(skb);
4922                 tcp_data_snd_check(sk);
4923                 return 0;
4924         }
4925
4926         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4927             tcp_paws_discard(sk, skb)) {
4928                 if (!th->rst) {
4929                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4930                         tcp_send_dupack(sk, skb);
4931                         goto discard;
4932                 }
4933                 /* Reset is accepted even if it did not pass PAWS. */
4934         }
4935
4936         /* step 1: check sequence number */
4937         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4938                 if (!th->rst)
4939                         tcp_send_dupack(sk, skb);
4940                 goto discard;
4941         }
4942
4943         /* step 2: check RST bit */
4944         if (th->rst) {
4945                 tcp_reset(sk);
4946                 goto discard;
4947         }
4948
4949         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4950
4951         /* step 3: check security and precedence [ignored] */
4952
4953         /*      step 4:
4954          *
4955          *      Check for a SYN in window.
4956          */
4957         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4958                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4959                 tcp_reset(sk);
4960                 return 1;
4961         }
4962
4963         /* step 5: check the ACK field */
4964         if (th->ack) {
4965                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4966
4967                 switch (sk->sk_state) {
4968                 case TCP_SYN_RECV:
4969                         if (acceptable) {
4970                                 tp->copied_seq = tp->rcv_nxt;
4971                                 smp_mb();
4972                                 tcp_set_state(sk, TCP_ESTABLISHED);
4973                                 sk->sk_state_change(sk);
4974
4975                                 /* Note, that this wakeup is only for marginal
4976                                  * crossed SYN case. Passively open sockets
4977                                  * are not waked up, because sk->sk_sleep ==
4978                                  * NULL and sk->sk_socket == NULL.
4979                                  */
4980                                 if (sk->sk_socket) {
4981                                         sk_wake_async(sk,0,POLL_OUT);
4982                                 }
4983
4984                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4985                                 tp->snd_wnd = ntohs(th->window) <<
4986                                               tp->rx_opt.snd_wscale;
4987                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4988                                             TCP_SKB_CB(skb)->seq);
4989
4990                                 /* tcp_ack considers this ACK as duplicate
4991                                  * and does not calculate rtt.
4992                                  * Fix it at least with timestamps.
4993                                  */
4994                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4995                                     !tp->srtt)
4996                                         tcp_ack_saw_tstamp(sk, 0);
4997
4998                                 if (tp->rx_opt.tstamp_ok)
4999                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5000
5001                                 /* Make sure socket is routed, for
5002                                  * correct metrics.
5003                                  */
5004                                 icsk->icsk_af_ops->rebuild_header(sk);
5005
5006                                 tcp_init_metrics(sk);
5007
5008                                 tcp_init_congestion_control(sk);
5009
5010                                 /* Prevent spurious tcp_cwnd_restart() on
5011                                  * first data packet.
5012                                  */
5013                                 tp->lsndtime = tcp_time_stamp;
5014
5015                                 tcp_mtup_init(sk);
5016                                 tcp_initialize_rcv_mss(sk);
5017                                 tcp_init_buffer_space(sk);
5018                                 tcp_fast_path_on(tp);
5019                         } else {
5020                                 return 1;
5021                         }
5022                         break;
5023
5024                 case TCP_FIN_WAIT1:
5025                         if (tp->snd_una == tp->write_seq) {
5026                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5027                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5028                                 dst_confirm(sk->sk_dst_cache);
5029
5030                                 if (!sock_flag(sk, SOCK_DEAD))
5031                                         /* Wake up lingering close() */
5032                                         sk->sk_state_change(sk);
5033                                 else {
5034                                         int tmo;
5035
5036                                         if (tp->linger2 < 0 ||
5037                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5038                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5039                                                 tcp_done(sk);
5040                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5041                                                 return 1;
5042                                         }
5043
5044                                         tmo = tcp_fin_time(sk);
5045                                         if (tmo > TCP_TIMEWAIT_LEN) {
5046                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5047                                         } else if (th->fin || sock_owned_by_user(sk)) {
5048                                                 /* Bad case. We could lose such FIN otherwise.
5049                                                  * It is not a big problem, but it looks confusing
5050                                                  * and not so rare event. We still can lose it now,
5051                                                  * if it spins in bh_lock_sock(), but it is really
5052                                                  * marginal case.
5053                                                  */
5054                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5055                                         } else {
5056                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5057                                                 goto discard;
5058                                         }
5059                                 }
5060                         }
5061                         break;
5062
5063                 case TCP_CLOSING:
5064                         if (tp->snd_una == tp->write_seq) {
5065                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5066                                 goto discard;
5067                         }
5068                         break;
5069
5070                 case TCP_LAST_ACK:
5071                         if (tp->snd_una == tp->write_seq) {
5072                                 tcp_update_metrics(sk);
5073                                 tcp_done(sk);
5074                                 goto discard;
5075                         }
5076                         break;
5077                 }
5078         } else
5079                 goto discard;
5080
5081         /* step 6: check the URG bit */
5082         tcp_urg(sk, skb, th);
5083
5084         /* step 7: process the segment text */
5085         switch (sk->sk_state) {
5086         case TCP_CLOSE_WAIT:
5087         case TCP_CLOSING:
5088         case TCP_LAST_ACK:
5089                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5090                         break;
5091         case TCP_FIN_WAIT1:
5092         case TCP_FIN_WAIT2:
5093                 /* RFC 793 says to queue data in these states,
5094                  * RFC 1122 says we MUST send a reset.
5095                  * BSD 4.4 also does reset.
5096                  */
5097                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5098                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5099                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5100                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5101                                 tcp_reset(sk);
5102                                 return 1;
5103                         }
5104                 }
5105                 /* Fall through */
5106         case TCP_ESTABLISHED:
5107                 tcp_data_queue(sk, skb);
5108                 queued = 1;
5109                 break;
5110         }
5111
5112         /* tcp_data could move socket to TIME-WAIT */
5113         if (sk->sk_state != TCP_CLOSE) {
5114                 tcp_data_snd_check(sk);
5115                 tcp_ack_snd_check(sk);
5116         }
5117
5118         if (!queued) {
5119 discard:
5120                 __kfree_skb(skb);
5121         }
5122         return 0;
5123 }
5124
5125 EXPORT_SYMBOL(sysctl_tcp_ecn);
5126 EXPORT_SYMBOL(sysctl_tcp_reordering);
5127 EXPORT_SYMBOL(tcp_parse_options);
5128 EXPORT_SYMBOL(tcp_rcv_established);
5129 EXPORT_SYMBOL(tcp_rcv_state_process);
5130 EXPORT_SYMBOL(tcp_initialize_rcv_mss);